copilot-custom-endpoint 3.4.4 → 3.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -22,17 +22,18 @@ That's it. No code, no servers to manage (unless the model specifically needs th
22
22
 
23
23
  | Model | Provider | Needs proxy? | Vision | Setup guide |
24
24
  | ---------------------------- | ---------- | ----------------------- | ------ | ----------------------------------------------------------------------------------------------------------------------------------------- |
25
+ | **Kimi K3.1 Pro** | Moonshot | **Yes** | ✅ | [Manual setup](docs/models/kimi.md) |
25
26
  | **Kimi K3** | Moonshot | **Yes** | ✅ | [Manual setup](docs/models/kimi.md) |
26
27
  | **MiMo V2.6 Pro** | Xiaomi | **Yes** | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=sdmapvstool.xiaomimimo-for-copilot) / [Manual setup](docs/models/mimo.md) |
27
28
  | **MiMo V2.6 Flash** | Xiaomi | **Yes** | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=sdmapvstool.xiaomimimo-for-copilot) / [Manual setup](docs/models/mimo.md) |
28
29
  | **MiMo V2.6 Pro UltraSpeed** | Xiaomi | **Yes** | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=sdmapvstool.xiaomimimo-for-copilot) / [Manual setup](docs/models/mimo.md) |
29
30
  | **Qwen 3.8 Max (0803)** | OpenRouter | Optional (experimental) | ✅ | [Snapshot/proxy setup](docs/models/qwen.md#optional-openrouter-snapshot-proxy) |
30
31
  | **Qwen 3.8 Max (0902)** | DashScope | Optional (recommended) | ✅ | [Manual setup](docs/models/qwen.md) |
32
+ | **MiniMax M3.1 Flash** | MiniMax | No | ✅ | [Manual setup](docs/models/minimax.md) |
31
33
  | **MiniMax M3** | MiniMax | No | ✅ | [Extension](https://github.com/tugudush/minimax-copilot) ★ / [Manual setup](docs/models/minimax.md) |
32
34
  | **GLM 5.3 Flash** | Z.ai | No | ✅ | [Manual setup](docs/models/glm.md) |
33
35
  | **GLM 5.3** | Z.ai | No | ❌ | [Manual setup](docs/models/glm.md) |
34
- | **DeepSeek V4.1 Flash** | DeepSeek | No (uses an extension) | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot) / [setup notes](docs/models/deepseek.md) |
35
- | **DeepSeek V4 Pro 0813** | DeepSeek | No (uses an extension) | ❌ | [Extension](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot) / [setup notes](docs/models/deepseek.md) |
36
+ | **DeepSeek V4.1 Flash 1017** | DeepSeek | No (uses an extension) | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot) / [setup notes](docs/models/deepseek.md) |
36
37
 
37
38
  ## Setup
38
39
 
@@ -121,7 +122,7 @@ It approximates the cost of one point on the AA Intelligence Index under the com
121
122
 
122
123
  For example, GLM 5.3 Flash is `(0.5 × $0.15) + (0.1 × $0.50) = $0.125`, shown as `~$0.13`. Cached-input discounts are not included. Actual costs vary with context length, output length, caching, pricing tiers, promotions, and provider billing. DeepSeek estimates use peak rates; official off-peak rates are half.
123
124
 
124
- The current AA values come from OpenRouter's model metadata and ranking card; **every pricing cell comes from first-party provider pricing or GitHub Copilot's official billing table, never OpenRouter.** The OpenRouter ranking card places unversioned Qwen3.8 Max at **53.4 (#3)**, while the exact 0902 API metadata row is **45.4**. Qwen Cloud PAYG pricing remains **$2 / $0.25 implicit cache / $6** for Qwen3.8 Max; its separate Token Plan is not the PAYG rate table. This snapshot records Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.7, GPT-6 Sol, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 as **Copilot-native** models, using current AA v4.3.2 composites. Gemini 4 Argon (released October 13, 2026) enters at **53** on Google's introductory pricing of **$2 / $0.10 implicit cache / $10** per 1M tokens, which doubles to **$4 / $0.20 / $20** on December 1, 2026. Claude Haiku 5.5 (released October 7, 2026) enters at **43** and, at **~$0.0023** per intelligence point, becomes the lowest cost per intelligence in the table; MiMo V2.6 Flash (**37.9**) remains the cheapest custom-endpoint option. Muse Spark 1.3 (Meta, released September 2, 2026) enters at **48** on Meta Model API pricing of **$1.25 / $0.15 cached input / $4.25** per 1M tokens; its day-one `muse-spark-1.3-contributor` SKU is excluded from this snapshot because Meta may train future models on those prompts and completions.
125
+ The current AA values come from OpenRouter's model metadata and ranking card; **every pricing cell comes from first-party provider pricing or GitHub Copilot's official billing table, never OpenRouter.** The OpenRouter ranking card places unversioned Qwen3.8 Max at **53.4 (#3)**, while the exact 0902 API metadata row is **45.4**. Qwen Cloud PAYG pricing remains **$2 / $0.25 implicit cache / $6** for Qwen3.8 Max; its separate Token Plan is not the PAYG rate table. This snapshot records Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.7, GPT-6 Sol, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 as **Copilot-native** models, using current AA v4.3.2 composites. Gemini 4 Argon (released October 13, 2026) enters at **53** on Google's introductory pricing of **$2 / $0.10 implicit cache / $10** per 1M tokens, which doubles to **$4 / $0.20 / $20** on December 1, 2026. Claude Haiku 5.5 (released October 7, 2026) enters at **43** and, at **~$0.0023** per intelligence point, becomes the lowest cost per intelligence in the table; MiMo V2.6 Flash (**37.9**) remains the cheapest custom-endpoint option. Muse Spark 1.3 (Meta, released September 2, 2026) enters at **48** on Meta Model API pricing of **$1.25 / $0.15 cached input / $4.25** per 1M tokens; its day-one `muse-spark-1.3-contributor` SKU is excluded from this snapshot because Meta may train future models on those prompts and completions. DeepSeek's **V4.1-Flash-1017** build (released October 17, 2026) replaces the September 10 build under the same `deepseek-flash` API ID: its composite rises from **39.5 to 45** at unchanged peak rates ($0.30 / $0.006 / $1.20), moving it to **~$0.0060** per intelligence point between GPT-5.6 Luna and MiMo V2.6 Pro. MiniMax's **MiniMax-M3.1-Flash** (announced October 19, 2026) enters at **43** on flat pricing of **$0.15 / $0.015 cache read / $0.60** per 1M tokens — no prompt-length tiers and no `service_tier` options — which puts it at **~$0.0031** per intelligence point between GLM 5.3 Flash and GPT-5.6 Luna. Artificial Analysis measures it at 96 tokens/s and very verbose (180M output tokens to run the index vs a 100M median), so real sessions can cost more than the modeled ~$0.14.
125
126
 
126
127
  | Model | Provider | Cost per intelligence | Intelligence Score | Est. session | Vision | Context window |
127
128
  | ---------------------------- | --------- | --------------------- | ------------------ | ------------ | ------ | -------------- |
@@ -129,19 +130,20 @@ The current AA values come from OpenRouter's model metadata and ranking card; **
129
130
  | **MiMo V2.6 Flash** | Xiaomi | **~$0.0026** | **37.9** | ~$0.10 | ✅ | 1M |
130
131
  | **GPT-6 Luna** | OpenAI | **~$0.0027** | **37.3** | ~$0.10 | ✅ | 1M |
131
132
  | **GLM 5.3 Flash** | Z.ai | **~$0.0030** | **41.8** | ~$0.13 | ✅ | 1M |
133
+ | **MiniMax M3.1 Flash** | MiniMax | **~$0.0031** | **43** | ~$0.14 | ✅ | 1M |
132
134
  | **GPT-5.6 Luna** | OpenAI | **~$0.0059** | **37.3** | ~$0.22 | ✅ | 1M |
135
+ | **DeepSeek V4.1 Flash 1017** | DeepSeek | **~$0.0060** | **45** | ~$0.27 | ✅ | 1M |
133
136
  | **MiMo V2.6 Pro** | Xiaomi | **~$0.0066** | **46.3** | ~$0.30 | ✅ | 1M |
134
- | **DeepSeek V4.1 Flash** | DeepSeek | **~$0.0068** | **39.5** | ~$0.27 | ✅ | 1M |
135
137
  | **MiniMax M3** | MiniMax | **~$0.0093** | **29.2** | ~$0.27 | ✅ | 1M |
136
138
  | **Gemini 3.8 Flash** | Google | **~$0.0183** | **40.9** | ~$0.75 | ✅ | 1M |
137
139
  | **Muse Spark 1.3** | Meta | **~$0.0219** | **48** | ~$1.05 | ✅ | 1M |
138
140
  | **GLM 5.3** | Z.ai | **~$0.0254** | **44.8** | ~$1.14 | ❌ | 1M |
139
- | **DeepSeek V4 Pro 0813** | DeepSeek | **~$0.0294** | **36.0** | ~$1.06 | ❌ | 1M |
140
141
  | **Qwen 3.8 Max (0803)** | DashScope | **~$0.0300** | **53.4** | ~$1.60 | ✅ | 1M |
141
142
  | **Grok 4.7** | xAI | **~$0.0345** | **46.4** | ~$1.60 | ✅ | 500K |
142
143
  | **Qwen 3.8 Max (0902)** | DashScope | **~$0.0352** | **45.4** | ~$1.60 | ✅ | 1M |
143
144
  | **Claude Sonnet 5.5** | Anthropic | **~$0.0357** | **56.0** | ~$2.00 | ✅ | 1M |
144
145
  | **Grok 4.6** | xAI | **~$0.0361** | **44.3** | ~$1.60 | ✅ | 500K |
146
+ | **Kimi K3.1 Pro** | Moonshot | **~$0.0363** | **55.1** | ~$2.00 | ✅ | 1M |
145
147
  | **Gemini 4 Argon** | Google | **~$0.0377** | **53** | ~$2.00 | ✅ | 1M |
146
148
  | **GPT-6.1 Sol** | OpenAI | **~$0.0386** | **51.8** | ~$2.00 | ✅ | 1M |
147
149
  | **GPT-6 Sol** | OpenAI | **~$0.0421** | **47.5** | ~$2.00 | ✅ | 1M |
@@ -156,7 +158,7 @@ The current AA values come from OpenRouter's model metadata and ranking card; **
156
158
  | **GPT-6 Astra** | OpenAI | **~$0.1898** | **52.7** | ~$10.00 | ✅ | 1M |
157
159
  | **Claude Fable 5** | Anthropic | **~$0.2016** | **49.6** | ~$10.00 | ✅ | 1M |
158
160
 
159
- > `MiMo V2.6 Pro UltraSpeed` uses the MiMo V2.6 Pro benchmark score because it is the same-quality latency tier; no independent composite is published. `MiniMax M3 Priority` is intentionally omitted from this snapshot because it is the same model with a priority service tier; the full pricing table documents its separate tier cost. `MiMo V2.6 Flash`'s composite (37.9) is OpenRouter's Artificial Analysis v4.3.2 value; Xiaomi's published launch benchmarks for it are recorded in [docs/benchmarks.md](docs/benchmarks.md).
161
+ > `MiMo V2.6 Pro UltraSpeed` uses the MiMo V2.6 Pro benchmark score because it is the same-quality latency tier; no independent composite is published. `MiniMax M3.1 Flash` is a separate model with its own flat rates; MiniMax's API rejects `service_tier` for it. `Kimi K3.1 Pro` supersedes K3 at a lower price ($2 / $0.20 / $10 vs $3 / $0.30 / $15 per 1M); K3 remains served. `MiMo V2.6 Flash`'s composite (37.9) is OpenRouter's Artificial Analysis v4.3.2 value; Xiaomi's published launch benchmarks for it are recorded in [docs/benchmarks.md](docs/benchmarks.md).
160
162
  > GPT-6 Sol, GPT-6.1 Sol, and Luna use standard rates up to 272K input tokens; requests above that threshold have higher input/cache and output rates. GitHub Copilot also bills cache writes separately. Plan access differs: GPT-6 Sol and GPT-6.1 Sol require Pro+ or higher, Luna and Claude Sonnet 5.5 are available from Pro, Claude Opus 5.5 requires Pro+ or higher, and Claude Haiku 5.5 is available on every paid plan (Pro through Enterprise). Gemini 4 Argon is Copilot-native for Pro+, Max, Business, and Enterprise (not Pro); Google's $2 / $0.10 implicit cache / $10 per 1M introductory pricing (output includes thinking tokens, implicit caching only) is guaranteed only through November 30, 2026 and doubles to $4 / $0.20 / $20 on December 1, 2026.
161
163
  > Claude Haiku 5.5 is the only current Claude model priced by prompt length: $0.10 / $0.01 cache read / $0.50 per 1M for prompts up to 100,000 tokens, and $0.50 / $0.05 / $2.50 above that. The table shows the sub-100K rates, which Anthropic says cover ~90% of Haiku-class requests.
162
164
  > Muse Spark 1.3 is closed and API-only on the Meta Model API (`muse-spark-1.3`); the 12–21× cheaper `muse-spark-1.3-contributor` SKU is excluded from this snapshot because Meta may use prompts and completions to train future models. Meta reports audio understanding is degraded versus 1.2.
@@ -2,7 +2,7 @@
2
2
 
3
3
  Here's a complete, real-world `chatLanguageModels.json` that combines **the `customendpoint` providers from the live `chatLanguageModels.json`**, plus the additional GLM models (`glm-5.3-flash`, `glm-5.3`, `glm-5.2`, `glm-5.1`) that are validated in [docs/models/glm.md](models/glm.md) but not present in every live config. Copy what you need, leave the rest out.
4
4
 
5
- > **⚠️ Reminder:** After adding models, you must also configure the **Utility Small Model** setting in VS Code. Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). Without this, custom-endpoint models may not function correctly. See the [main README](../README.md#4-configure-the-utility-small-model) for details.
5
+ > **⚠️ Reminder:** After adding models, you must also configure the **Utility Small Model** setting in VS Code. Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). Without this, custom-endpoint models may not function correctly. See the [main README](../README.md#4-configure-the-utility-small-model) for details.
6
6
  >
7
7
  > **Note:** The `apiKey` fields are left as empty strings — set them via the **Chat: Manage Language Models** UI (Command Palette → right-click provider group → **Update API Key**). After you set a key via the UI, VS Code replaces the empty string with a `${input:chat.lm.secret.<id>}` secret reference.
8
8
  >
@@ -34,6 +34,20 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
34
34
  "apiKey": "",
35
35
  "apiType": "chat-completions",
36
36
  "models": [
37
+ {
38
+ "id": "kimi-k3.1-pro",
39
+ "name": "Kimi K3.1 Pro (vision)",
40
+ "url": "http://127.0.0.1:3457/v1/chat/completions",
41
+ "requestBody": {
42
+ "temperature": 1,
43
+ "max_tokens": 8192
44
+ },
45
+ "toolCalling": true,
46
+ "vision": true,
47
+ "streaming": true,
48
+ "maxInputTokens": 1000000,
49
+ "maxOutputTokens": 131072
50
+ },
37
51
  {
38
52
  "id": "kimi-k3",
39
53
  "name": "Kimi K3 (vision)",
@@ -106,6 +120,23 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
106
120
  "apiKey": "",
107
121
  "apiType": "chat-completions",
108
122
  "models": [
123
+ {
124
+ "id": "MiniMax-M3.1-Flash",
125
+ "name": "MiniMax M3.1 Flash (vision)",
126
+ "url": "https://api.minimax.io/v1/chat/completions",
127
+ "toolCalling": true,
128
+ "vision": true,
129
+ "streaming": true,
130
+ "maxInputTokens": 1048576,
131
+ "maxOutputTokens": 131072,
132
+ "requestBody": {
133
+ "thinking": { "type": "adaptive" },
134
+ "reasoning_split": true,
135
+ "reasoning_effort": "high",
136
+ "temperature": 1,
137
+ "top_p": 0.95
138
+ }
139
+ },
109
140
  {
110
141
  "id": "MiniMax-M3",
111
142
  "name": "MiniMax M3 (vision)",
@@ -174,7 +205,7 @@ If you only need one provider, jump straight to its setup guide:
174
205
  - [Kimi K3 / K2.6 / K2.7 Code](kimi.md)
175
206
  - [Qwen 3.8 Max](qwen.md)
176
207
  - [Xiaomi MiMo (V2.6 Pro / Flash / Pro UltraSpeed)](mimo.md)
177
- - [MiniMax M3 (Standard)](minimax.md) and [MiniMax M3 (Priority)](minimax.md#5-m3-priority-tier-optional)
208
+ - [MiniMax M3.1 Flash / M3](minimax.md)
178
209
  - [GLM (5.3 Flash / 5.3 / 5.2 / 5.1)](glm.md)
179
210
 
180
- > **DeepSeek V4 Pro 0813 / V4 Flash 0731** use the [DeepSeek V4 for Copilot Chat](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot) extension. They appear in `chatLanguageModels.json` as `vendor: "deepseek"` (not as a `customendpoint` provider) and are configured via the extension's settings block.
211
+ > **DeepSeek V4.1 Flash 1017** uses the [DeepSeek for Copilot Chat extension](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot). It appears in `chatLanguageModels.json` as `vendor: "deepseek"` (not as a `customendpoint` provider) and is configured via the extension's settings block.
@@ -1,58 +1,44 @@
1
- # DeepSeek V4 for Copilot Chat
1
+ # DeepSeek V4.1 Flash for Copilot Chat
2
2
 
3
- DeepSeek V4 Pro and V4.1 Flash are provided by the [DeepSeek V4 for Copilot Chat](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot) extension. They do not use this repository's `chatLanguageModels.json` custom-endpoint configuration.
3
+ DeepSeek V4.1 Flash is provided by the [DeepSeek for Copilot Chat extension](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot). It does not use this repository's `chatLanguageModels.json` custom-endpoint configuration.
4
4
 
5
- ## Current builds
5
+ ## Current build
6
6
 
7
- The current API model IDs and builds are:
7
+ The current API model ID and build are:
8
8
 
9
- | API model | Current version | Update | Context | Modality | Parameters |
10
- | ----------------- | -------------------- | ------------------------------------- | ------- | -------------------------- | --------------------------------------------------------------- |
11
- | `deepseek-flash` | DeepSeek-V4.1-Flash | September 10, 2026; current API build | 1M | Text + image in / text out | Causal Encoder-Decoder sparse MoE; 8B input / 16B output active |
12
- | `deepseek-v4-pro` | DeepSeek-V4-Pro-0813 | August 13, 2026; GA release | 1M | Text in / text out | 1.6T total / 49B active |
9
+ | API model | Current version | Update | Context | Modality | Parameters |
10
+ | ---------------- | ------------------------ | ----------------------------------- | ------- | -------------------------- | --------------------------------------------------------------- |
11
+ | `deepseek-flash` | DeepSeek-V4.1-Flash-1017 | October 17, 2026; current API build | 1M | Text + image in / text out | Causal Encoder-Decoder sparse MoE; 8B input / 16B output active |
13
12
 
14
- The [DeepSeek API pricing page](https://api-docs.deepseek.com/quick_start/pricing) identifies `deepseek-flash` as the current Flash model and says the legacy `deepseek-v4-flash` names remain accepted but are retired. Do not add the version suffix to the API request or the extension override.
13
+ The [DeepSeek API pricing page](https://api-docs.deepseek.com/quick_start/pricing) identifies `deepseek-flash` as the current Flash model and reports its model version as **DeepSeek-V4.1-Flash-1017**. Do not add the version suffix to the API request or the extension override.
15
14
 
16
- ## Model ID overrides
17
-
18
- The extension's `deepseek-copilot.modelIdOverrides` setting changes the upstream API model IDs while keeping the normal DeepSeek entries in the Copilot model picker. The current official DeepSeek Chat Completions API accepts `deepseek-flash` and `deepseek-v4-pro`, so map the extension's logical Flash entry to the canonical current ID:
15
+ DeepSeek's [October 17 release note](https://api-docs.deepseek.com/news/news261017) describes 1017 as a Pareto improvement over the September 10 build: stronger benchmarks, faster task completion, and shorter outputs. The model supports both non-thinking and thinking (default) modes, keeps the 1M context window and vision input, and publishes open weights on [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) (552B total / 16B active parameters, MIT license).
19
16
 
20
- ```jsonc
21
- {
22
- "deepseek-copilot.modelIdOverrides": {
23
- "deepseek-v4-flash": "deepseek-flash",
24
- "deepseek-v4-pro": "deepseek-v4-pro"
25
- }
26
- }
27
- ```
28
-
29
- > **Warning:** Do **not** use the human-readable version labels as model IDs. `DeepSeek-V4.1-Flash` and `DeepSeek-V4-Pro-0813` are version strings, not valid `model` fields. Use `deepseek-flash` for current Flash requests and `deepseek-v4-pro` for Pro. Sending a version-suffixed ID to `https://api.deepseek.com` can return **HTTP 400 "Invalid request body format"**.
17
+ ## Model ID overrides
30
18
 
31
- Only add an override when the configured `deepseek-copilot.baseUrl` expects a different model name (e.g., a third-party proxy). The left-hand keys are the extension's logical model keys; they are not arbitrary picker labels. The right-hand values are sent verbatim as the provider's `model` field.
19
+ The extension's `deepseek-copilot.modelIdOverrides` setting changes the upstream API model ID while keeping the normal DeepSeek entry in the Copilot model picker. The current official DeepSeek Chat Completions API accepts `deepseek-flash`, so map the extension's logical Flash entry to that canonical ID.
32
20
 
33
- For Flash or Pro, select the corresponding **DeepSeek V4** entry in Copilot Chat after reloading the window. The picker may still display the family name without the build suffix; that is expected. The override is not a model registration or a way to add a new picker entry.
21
+ > **Warning:** Do **not** use the human-readable version label as a model ID. `DeepSeek-V4.1-Flash-1017` is a version string, not a valid `model` field; use `deepseek-flash`. Sending a version-suffixed ID to `https://api.deepseek.com` can return **HTTP 400 "Invalid request body format"**.
34
22
 
35
- ## Updated benchmarks
23
+ Only add an override when the configured `deepseek-copilot.baseUrl` expects a different model name (e.g., a third-party proxy). The left-hand key is the extension's logical model key; it is not an arbitrary picker label. The right-hand value is sent verbatim as the provider's `model` field.
36
24
 
37
- DeepSeek's official August updates report the following agent-focused results for the retired 0731 Flash release and the GA Pro release:
25
+ Select the **DeepSeek V4.1 Flash** entry in Copilot Chat after reloading the window. The picker may still display the family name without the build suffix; that is expected. The override is not a model registration or a way to add a new picker entry.
38
26
 
39
- - **Historical Flash 0731:** Terminal-Bench 2.1 **82.7**, NL2Repo **54.2**, Cybergym **76.7**, DeepSWE **54.4**, Toolathlon Verified **70.3**, Agents' Last Exam **25.2**, AutomationBench (Public) **25.1**, DSBench-FullStack **68.7**, and DSBench-Hard **59.6**.
40
- - **Pro 0813:** HLE without/with tools **42.7/60.0**, Terminal-Bench 2.1 **87.9**, NL2Repo **61.5**, Cybergym **83.3**, DeepSWE **62.7**, Toolathlon Verified **74.1**, Agents' Last Exam **25.7**, AutomationBench (Public) **31.8**, DSBench-FullStack **71.1**, and DSBench-Hard **67.2**.
27
+ ## Benchmarks
41
28
 
42
- The current OpenRouter metadata reports **V4.1 Flash 39.5 AA Intelligence** and **Pro 36.3 AA Intelligence**. Current Flash Coding and Agentic composites were not exposed in the checked metadata. The older 0731 snapshot reported **51.8 AA Intelligence / 69.1 Coding / 48.4 Agentic**, while Pro's current metadata reports **36.3 AA Intelligence / 68.8 Coding / 49.6 Agentic**. AA values are independent composite scores; DeepSeek's agent results above are vendor-reported and should not be treated as the same metric.
29
+ The current Artificial Analysis v4.3.2 composite for **V4.1 Flash 1017** is **45**, up from **39.5** for the superseded September 10 build (whose AA page is now marked deprecated). AA measures 1017 at 216 tokens/s and **$0.26 per Intelligence Index task**, and it generated **210M output tokens** to run the index versus a 140M median (the September build used 250M), so it remains fast but very verbose. Current Coding and Agentic composites were not exposed in the checked metadata. DeepSeek's [release note](https://api-docs.deepseek.com/news/news261017) reports stronger benchmark performance, faster task completion, and shorter outputs than the September build; those vendor results are not directly comparable to the AA composite.
43
30
 
44
- Sources: [DeepSeek pricing and model page](https://api-docs.deepseek.com/quick_start/pricing), [Flash benchmark snapshot](https://openrouter.ai/deepseek/deepseek-v4.1-flash), [Pro benchmark snapshot](https://openrouter.ai/deepseek/deepseek-v4-pro-0813), [Flash AA page](https://artificialanalysis.ai/models/deepseek-v4-1-flash), and [Pro AA page](https://artificialanalysis.ai/models/deepseek-v4-pro).
31
+ Sources: [DeepSeek pricing and model page](https://api-docs.deepseek.com/quick_start/pricing), [DeepSeek 1017 release note](https://api-docs.deepseek.com/news/news261017), [1017 benchmark snapshot](https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017), and [1017 AA page](https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017).
45
32
 
46
33
  ## Current API pricing
47
34
 
48
- DeepSeek switched V4 pricing to peak/off-peak rates at **16:00 UTC on August 16, 2026**. Prices below are USD per 1M tokens. Peak hours are **01:00-04:00 and 06:00-10:00 UTC**; off-peak rates are half the peak rates.
35
+ DeepSeek switched to peak/off-peak rates at **16:00 UTC on August 16, 2026**. Prices below are USD per 1M tokens. Peak hours are **01:00-04:00 and 06:00-10:00 UTC**; off-peak rates are half the peak rates.
49
36
 
50
- | API model | Cache hit | Cache miss | Output |
51
- | ----------------- | ----------------------------- | --------------------------- | --------------------------- |
52
- | `deepseek-flash` | $0.003 off-peak / $0.006 peak | $0.15 off-peak / $0.30 peak | $0.60 off-peak / $1.20 peak |
53
- | `deepseek-v4-pro` | $0.022 off-peak / $0.044 peak | $0.66 off-peak / $1.32 peak | $1.98 off-peak / $3.96 peak |
37
+ | API model | Cache hit | Cache miss | Output |
38
+ | ---------------- | ----------------------------- | --------------------------- | --------------------------- |
39
+ | `deepseek-flash` | $0.003 off-peak / $0.006 peak | $0.15 off-peak / $0.30 peak | $0.60 off-peak / $1.20 peak |
54
40
 
55
- The legacy `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` names remain accepted but are retired aliases for V4.1 Flash. See the [official Models & Pricing page](https://api-docs.deepseek.com/quick_start/pricing) for the live rates and concurrency limits.
41
+ The 1017 build keeps these rates unchanged. See the [official Models & Pricing page](https://api-docs.deepseek.com/quick_start/pricing) for the live rates and concurrency limits.
56
42
 
57
43
  ## Tool-list stabilization (`experimental.stabilizeToolList`)
58
44
 
@@ -62,13 +48,13 @@ The legacy `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` names remain a
62
48
  - **When off**, the preflight pass is skipped entirely.
63
49
  - **Plain chat (no tools) never triggers the flag** — it is inert and has no effect (and no cost) on non-tool requests. It neither causes nor fixes plain-chat 400s such as the invalid model-ID case above.
64
50
 
65
- Recommendation: leave it `true` if you use **agent mode or MCP tools** with DeepSeek V4; it is harmless to keep for plain-chat-only use, and safe to remove if you never invoke tools.
51
+ Recommendation: leave it `true` if you use **agent mode or MCP tools** with DeepSeek V4.1 Flash; it is harmless to keep for plain-chat-only use, and safe to remove if you never invoke tools.
66
52
 
67
53
  ## Troubleshooting
68
54
 
69
55
  1. Confirm the extension is installed and update it if the setting is not recognized. The setting is owned by `Vizards.deepseek-v4-for-copilot`.
70
- 2. Use the exact setting name and logical key:
71
- `deepseek-copilot.modelIdOverrides` → `deepseek-flash`.
56
+ 2. Use the exact setting name and point the Flash entry at `deepseek-flash`:
57
+ `deepseek-copilot.modelIdOverrides`.
72
58
  3. Reload VS Code after changing the setting (`Developer: Reload Window`).
73
59
  4. Verify the extension's `deepseek-copilot.baseUrl` points at the compatible provider or proxy. An override changes only the model ID; it does not change the endpoint, authentication, request format, or model capabilities.
74
60
  5. For HTTP 400 errors during agent mode or MCP/tool use, enable the extension's tool-list compatibility option:
@@ -44,7 +44,7 @@
44
44
  2. **Generate an API key** on the API Keys page of the platform.
45
45
  3. **Edit `chatLanguageModels.json`** — add the GLM block below.
46
46
  4. **Set the API key** via Command Palette → **Chat: Manage Language Models** → right-click **GLM** → **Update API Key**.
47
- 5. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
47
+ 5. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
48
48
  6. **Restart VS Code** and select a GLM model from the picker.
49
49
 
50
50
  ## Setup
@@ -1,8 +1,10 @@
1
1
  # Kimi (Moonshot) — VS Code Custom Endpoint Setup Guide
2
2
 
3
- > **TL;DR:** Kimi K3 requires the local proxy. It is always-thinking and uses `reasoning_effort` (not the `thinking` parameter) while the proxy enforces Kimi's sampling constraints. Direct VS Code → Moonshot is not viable.
3
+ > **TL;DR:** The Kimi K3 family requires the local proxy. These models are always-thinking and use `reasoning_effort` (not the `thinking` parameter) while the proxy enforces Kimi's fixed sampling constraints. Direct VS Code → Moonshot is not viable.
4
4
  >
5
- > **🆕 Kimi K3** (released July 16, 2026) is Kimi's flagship — 2.8T parameters, 1M context, AA Intelligence Index **59.7** (#4 overall, between GPT-5.6 Sol and Qwen 3.8 Max). Open-source weights coming by July 27, 2026.
5
+ > **🆕 Kimi K3.1 Pro** (released October 23, 2026) is the current flagship — a post-trained successor to K3 (3T class, 1M context) with AA Intelligence Index **55.1**, ahead of Claude Fable 5.1 and behind Claude Sonnet 5.5 in our roster. Always-thinking with `reasoning_effort` `low` / `high` / `max` (default `max`). Open weights promised by November 4, 2026.
6
+ >
7
+ > **Previous:** **Kimi K3** (July 16, 2026) — 2.8T parameters, 1M context, AA Index **43.6**. Still served and still supported by the proxy; use K3.1 Pro unless you specifically need K3.
6
8
  >
7
9
  > **🆕 Alternative:** On the **Kimi Coding tier** (`kimi.com/code/console`), the [**Moonshot LM Provider**](https://marketplace.visualstudio.com/items?itemName=DenizhanDaklr.kimi-lm-provider) extension offers zero-proxy setup with reasoning visibility. **Not compatible with Pay-as-You-Go** keys from `platform.kimi.ai`. See [the extension comparison](../research/kimi-vscode-extension.md).
8
10
 
@@ -22,19 +24,21 @@
22
24
 
23
25
  ### Models
24
26
 
25
- | Model | Vision | Context | Max output | Notes |
26
- | --------- | ------ | ------- | ---------- | -------------------------------------------------------------------------------------------------------- |
27
- | `kimi-k3` | ✅ Yes | 1M | 131072 | **Flagship.** Always-thinking, uses `reasoning_effort` (not `thinking`). 2.8T params. AA Index **59.7**. |
27
+ | Model | Vision | Context | Max output | Notes |
28
+ | ------------------------- | ------ | ------- | --------------- | ---------------------------------------------------------------------------------------------------------------------- |
29
+ | `kimi-k3.1-pro` | ✅ Yes | 1M | 131072 (1M max) | **Flagship.** Post-trained K3, 3T class. Always-thinking, uses `reasoning_effort` (not `thinking`). AA Index **55.1**. |
30
+ | `kimi-k3.1-pro-highspeed` | ✅ Yes | 1M | 131072 | 2–3× faster SKU of K3.1 Pro; same capabilities at a higher per-token price. |
31
+ | `kimi-k3` | ✅ Yes | 1M | 131072 | Previous flagship. 2.8T params. AA Index **43.6**. |
28
32
 
29
- > This guide documents the K3 setup. Deprecated K2 ids (`kimi-k2-0905-preview`, `kimi-k2-turbo-preview`, `kimi-k2-thinking`, etc.) were discontinued May 25, 2026.
33
+ > This guide documents the K3.1 Pro setup. K3.1 Pro and K3 share the same proxy and request shape, so the snippet below works for both — change only the `id`. Deprecated K2 ids (`kimi-k2-0905-preview`, `kimi-k2-turbo-preview`, `kimi-k2-thinking`, etc.) were discontinued May 25, 2026, and Moonshot plans to retire the remaining K2 series by the end of December 2026.
30
34
 
31
35
  ## Quick Start
32
36
 
33
37
  1. **Start the proxy** — `npm run proxy:kimi` (or `npx copilot-custom-endpoint kimi` standalone).
34
38
  2. **Edit `chatLanguageModels.json`** — add the Kimi block below.
35
39
  3. **Set your Moonshot API key** via Command Palette → **Chat: Manage Language Models**.
36
- 4. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
37
- 5. **Restart VS Code** and pick "Kimi K3".
40
+ 4. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
41
+ 5. **Restart VS Code** and pick "Kimi K3.1 Pro".
38
42
 
39
43
  ## Setup
40
44
 
@@ -53,9 +57,20 @@
53
57
  "apiKey": "",
54
58
  "apiType": "chat-completions",
55
59
  "models": [
60
+ {
61
+ "id": "kimi-k3.1-pro",
62
+ "name": "Kimi K3.1 Pro (Custom Endpoint)",
63
+ "url": "http://127.0.0.1:3457/v1/chat/completions",
64
+ "requestBody": { "temperature": 1, "max_tokens": 8192 },
65
+ "toolCalling": true,
66
+ "vision": true,
67
+ "streaming": true,
68
+ "maxInputTokens": 1000000,
69
+ "maxOutputTokens": 131072
70
+ },
56
71
  {
57
72
  "id": "kimi-k3",
58
- "name": "Kimi K3 (vision)",
73
+ "name": "Kimi K3 (Custom Endpoint)",
59
74
  "url": "http://127.0.0.1:3457/v1/chat/completions",
60
75
  "requestBody": { "temperature": 1, "max_tokens": 8192 },
61
76
  "toolCalling": true,
@@ -68,7 +83,9 @@
68
83
  }
69
84
  ```
70
85
 
71
- > **K3 note:** `max_tokens` is set to **8192** as a safe starting point. K3 is always-thinking and very verbose (130M tokens generated on the AA Intelligence Index evaluation). The API default for `max_completion_tokens` is 131072 and can go up to 1048576, but reasoning tokens inflate response size — monitor your usage and adjust upward if needed. Unlike K2.7, K3 uses `reasoning_effort` (not `thinking`); the proxy handles this automatically.
86
+ > **K3.1 Pro note:** `max_tokens` is set to **8192** as a safe starting point. K3.1 Pro is always-thinking and very verbose (150M tokens generated on the AA Intelligence Index evaluation versus an 87M median). The API default for `max_completion_tokens` is 131072 and can go up to 1048576, but reasoning tokens inflate response size — monitor your usage and adjust upward if needed. K3.1 Pro uses `reasoning_effort` (`low` / `high` / `max`, default `max`), not the K2.x `thinking` parameter; add `"reasoning_effort": "low"` to `requestBody` if answers take too long. The proxy forwards the value untouched.
87
+ >
88
+ > Both models are in the snippet above — delete the entry you don't plan to use. K3 accepts the same `requestBody`; it only supports `reasoning_effort: "max"`.
72
89
 
73
90
  ### 2. API key
74
91
 
@@ -112,16 +129,17 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
112
129
  - Forwards your `Authorization` header upstream.
113
130
  - Rewrites plain-chat requests to `temperature: 1`, `top_p: 0.95`.
114
131
  - For **K2.5 / K2.6**: rewrites tool-enabled requests to `thinking: { type: "disabled" }`, `temperature: 0.6`, `top_p: 0.95`.
115
- - For **K3**: keeps thinking enabled (K3 is always-thinking and uses `reasoning_effort`, not `thinking`); rewrites to `temperature: 1`, `top_p: 0.95`. Does **not** inject a `thinking` block — K3 rejects it.
132
+ - For **K3 / K3.1 Pro**: keeps thinking enabled (both are always-thinking and use `reasoning_effort`, not `thinking`); rewrites to `temperature: 1`, `top_p: 0.95`. Does **not** inject a `thinking` block — the K3 family rejects it. Every `kimi-k3*` slug is handled this way, including `kimi-k3.1-pro` and `kimi-k3.1-pro-highspeed`.
116
133
  - Preserves streaming responses (SSE).
117
134
  - Writes redacted request summaries to `debug_log/kimi-proxy.ndjson`.
118
135
 
119
136
  ## Notes
120
137
 
121
- - **Kimi K3** (released July 16, 2026) is Kimi's new flagship. 2.8T params, 1M context, always-thinking. Uses `reasoning_effort` (currently only `max`) — do **not** send the K2.x `thinking` parameter. `tool_choice` supports `auto`, `required`, and named tools. `max_completion_tokens` defaults to 131072 and can go up to 1048576. Fixed sampling: `temperature=1`, `top_p=0.95`. Vision supports text + image + video (via `ms://<file-id>` URLs). See the [K3 quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart).
138
+ - **Kimi K3.1 Pro** (released October 23, 2026) is the current flagship: a post-trained K3 with substantially larger post-training compute, 3T class, 1M context, open weights promised by November 4, 2026. Always-thinking — chain-of-thought **cannot** be disabled; tune depth with `reasoning_effort` (`low` / `high` / `max`, default `max`). Do **not** send the K2.x `thinking` parameter. `tool_choice` supports `auto`, `required`, and named tools. `max_completion_tokens` defaults to 131072 and can go up to 1048576. Sampling is fixed at `temperature=1`, `top_p=0.95`, `n=1`, `presence_penalty=0`, `frequency_penalty=0` — the proxy's enforced `1` / `0.95` are the only accepted values. Vision supports text + image plus video via `ms://<file-id>` URLs; public image URLs are rejected, so pass base64 or an uploaded file id. See the [K3.1 Pro quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart).
139
+ - **Kimi K3** (released July 16, 2026) remains served but is superseded by K3.1 Pro: 2.8T params, 1M context, always-thinking with `reasoning_effort` `max` only. It shares K3.1 Pro's request shape, so switching between them needs only a new `id` — no proxy or `requestBody` change. See the [K3 quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart).
122
140
  - **Kimi Platform rebranding:** the Kimi Platform (formerly Moonshot) rebranded to `platform.kimi.ai` in 2026 — `api.moonshot.ai` still resolves and the proxy targets it unchanged.
123
- - **`tool_choice`:** K3 supports `auto`, `required`, and named tools. The proxy preserves the value and only enforces temperature/top_p.
124
- - **K3 reasoning:** K3 uses `reasoning_effort` (currently only `max`) at the top level of the request body. Do not send the legacy `thinking` parameter — K3 rejects it.
141
+ - **`tool_choice`:** K3 and K3.1 Pro support `auto`, `required`, and named tools. The proxy preserves the value and only enforces temperature/top_p.
142
+ - **Do not send `thinking`:** the K3 family rejects the legacy K2.x `thinking` parameter — use `reasoning_effort` at the top level of the request body instead.
125
143
 
126
144
  ## Troubleshooting
127
145
 
@@ -132,6 +150,7 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
132
150
  | `invalid top_p: only 0.95 is allowed` | Direct path without proxy | Use the proxy |
133
151
  | `thinking is enabled but reasoning_content is missing` | Tool turn with thinking on | Verify `KIMI_PROXY_DISABLE_THINKING_WITH_TOOLS=1` |
134
152
  | `unsupported parameter: thinking` with K3 | K3 doesn't accept `thinking` | Use `reasoning_effort` instead; proxy detects K3 |
153
+ | Reasoning can't be turned off | K3.1 Pro always thinks | Set `reasoning_effort` to `low` instead |
135
154
  | Model not in VS Code picker | Config not reloaded | Restart VS Code |
136
155
  | `tool_choice=required` rejected (K2.x) | Model limitation | Use `auto` only for K2.x; K3 supports `required` |
137
156
 
@@ -139,8 +158,9 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
139
158
 
140
159
  For the cross-provider comparison, see [docs/pricing.md](../pricing.md). Moonshot direct platform rates:
141
160
 
142
- | Model | Input (cache miss) | Cached input | Output |
143
- | --------- | ------------------ | ------------ | ----------- |
144
- | `kimi-k3` | $3.00 / 1M | $0.30 / 1M | $15.00 / 1M |
161
+ | Model | Input (cache miss) | Cached input | Output |
162
+ | --------------- | ------------------ | ------------ | ----------- |
163
+ | `kimi-k3.1-pro` | $2.00 / 1M | $0.20 / 1M | $10.00 / 1M |
164
+ | `kimi-k3` | $3.00 / 1M | $0.30 / 1M | $15.00 / 1M |
145
165
 
146
- > K3 pricing is flat (no tiering by context length).
166
+ > K3 and K3.1 Pro pricing is flat (no tiering by context length), and caching is automatic for repeated prefixes — a 90% discount on cached input. Cache writes are billed separately by TTL (5 minutes / 1 hour) on both models: K3's write rates are $3.00/M (5 min) and $6.00/M (1 h). The `kimi-k3.1-pro-highspeed` SKU costs more than K3.1 Pro. Check the [official pricing table](https://platform.kimi.ai/docs/pricing/chat) for current write rates and the Highspeed rate.
@@ -57,7 +57,7 @@ Simplest no-proxy approach — but thinking is always off, so you never see mode
57
57
 
58
58
  1. **Edit `chatLanguageModels.json`** — add the MiMo block(s) below.
59
59
  2. **Set your `MIMO_API_KEY`** via Command Palette → **Chat: Manage Language Models**.
60
- 3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
60
+ 3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
61
61
  4. **Restart VS Code** and pick one of the V2.6 entries.
62
62
 
63
63
  ### Alternative: With optional proxy (dynamic thinking)
@@ -1,8 +1,6 @@
1
1
  # MiniMax — VS Code Custom Endpoint Setup Guide
2
2
 
3
- > **TL;DR:** MiniMax-M3 works directly — no proxy needed. Use `thinking: { type: "adaptive" }` + `reasoning_split: true` in `requestBody` so the model can reason and the response arrives in a clean OpenAI format (`reasoning_details` field, separate from `content`). **Important:** `thinking: { type: "disabled" }` is **not** a hard override — the model still reasons internally and emits `<think>` tags / `reasoning_content` regardless.
4
- >
5
- > Want faster responses and fewer admission failures during peak hours? Add `"service_tier": "priority"` to the single `MiniMax-M3` entry's `requestBody` — same model, **1.5× cost** for **priority admission**. (You can only toggle `service_tier` on the existing entry; registering a second picker entry with a different `id` is not valid on the custom-endpoint path because VS Code forwards the `id` as the upstream `model` parameter.) See [M3 Priority tier](#4-m3-priority-tier-optional) below.
3
+ > **TL;DR:** **MiniMax-M3.1-Flash** (released October 19, 2026) is the current M-series Flash model and needs no proxy. Use `thinking: { "type": "adaptive" }` + `reasoning_split: true` + `reasoning_effort` in `requestBody` so the model can reason and the response arrives in a clean OpenAI format (`reasoning_content`, kept separate from `content`). **M3.1 Flash cannot disable thinking** — `thinking: { "type": "disabled" }` and `reasoning_effort: "none"` both return HTTP 400 (`requires adaptive thinking`, code 2013) — so tune depth with `reasoning_effort` (`low` / `medium` / `high` / `xhigh` / `max`, default `max`). `MiniMax-M3` uses the same request shape, still accepts `thinking: { "type": "disabled" }` as a soft hint.
6
4
  >
7
5
  > The same `url`, `model id`, and `requestBody` work for **both** Pay-as-You-Go (account-balance billing) and Token Plan (monthly/annual subscription) — only the API key in the secret field changes.
8
6
  >
@@ -15,31 +13,32 @@
15
13
 
16
14
  ## At a Glance
17
15
 
18
- | Field | Value |
19
- | ------------------------ | ----------------------------------------------------------------------------------- |
20
- | Mode | **Direct** (no proxy) |
21
- | Billing | **Pay-as-You-Go** _or_ **Token Plan subscription** (same config) |
22
- | Vision | ✅ Yes (image + video) |
23
- | Tool calling | ✅ Yes |
24
- | Context | 1M (guaranteed 512K) |
25
- | Max output | 131072 |
26
- | Required `requestBody` | `thinking: { type: "adaptive" }, reasoning_split: true` |
27
- | Endpoint (international) | `https://api.minimax.io/v1/chat/completions` |
28
- | Endpoint (China) | `https://api.minimaxi.com/v1/chat/completions` |
29
- | Auth | `Authorization: Bearer $MINIMAX_API_KEY` (or Subscription Key, see [Setup](#setup)) |
16
+ | Field | Value |
17
+ | ------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------- |
18
+ | Mode | **Direct** (no proxy) |
19
+ | Billing | **Pay-as-You-Go** _or_ **Token Plan subscription** (same config) |
20
+ | Vision | ✅ Yes (image + video) |
21
+ | Tool calling | ✅ Yes |
22
+ | Context | 1M (M3 guarantees 512K) |
23
+ | Max output | 131072 recommended (524288 max) |
24
+ | Required `requestBody` | M3.1 Flash: `thinking: { type: "adaptive" }, reasoning_split: true, reasoning_effort: "high"` · M3: `thinking: { type: "adaptive" }, reasoning_split: true` |
25
+ | Endpoint (international) | `https://api.minimax.io/v1/chat/completions` |
26
+ | Endpoint (China) | `https://api.minimaxi.com/v1/chat/completions` |
27
+ | Auth | `Authorization: Bearer $MINIMAX_API_KEY` (or Subscription Key, see [Setup](#setup)) |
30
28
 
31
29
  ### Models
32
30
 
33
- | Model | Vision | Context | Max output | Notes |
34
- | ------------ | ------ | ------- | ---------- | -------------------------------------------------------------------------------------- |
35
- | `MiniMax-M3` | ✅ Yes | 1M | 131072 | Multimodal frontier coding model; adaptive reasoning; PAYG and Token Plan share config |
31
+ | Model | Vision | Context | Max output | Notes |
32
+ | -------------------- | ------ | ------- | ------------------------------- | ------------------------------------------------------------------------------------------------------ |
33
+ | `MiniMax-M3.1-Flash` | ✅ Yes | 1M | 131072 recommended (524288 max) | Announced October 19, 2026; tunable thinking depth via `reasoning_effort`; open weights promised later |
34
+ | `MiniMax-M3` | ✅ Yes | 1M | 131072 | Multimodal frontier coding model; adaptive reasoning; PAYG and Token Plan share config |
36
35
 
37
36
  ## Quick Start
38
37
 
39
38
  1. **Edit `chatLanguageModels.json`** — add the MiniMax block below.
40
39
  2. **Set your `MINIMAX_API_KEY`** via Command Palette → **Chat: Manage Language Models**.
41
- 3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
42
- 4. **Restart VS Code** and pick "MiniMax M3" in the chat picker.
40
+ 3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
41
+ 4. **Restart VS Code** and pick "MiniMax M3.1 Flash (vision)" (or "MiniMax M3") in the chat picker.
43
42
 
44
43
  ## Setup
45
44
 
@@ -67,6 +66,23 @@
67
66
  "apiKey": "",
68
67
  "apiType": "chat-completions",
69
68
  "models": [
69
+ {
70
+ "id": "MiniMax-M3.1-Flash",
71
+ "name": "MiniMax M3.1 Flash (vision)",
72
+ "url": "https://api.minimax.io/v1/chat/completions",
73
+ "toolCalling": true,
74
+ "vision": true,
75
+ "streaming": true,
76
+ "maxInputTokens": 1048576,
77
+ "maxOutputTokens": 131072,
78
+ "requestBody": {
79
+ "thinking": { "type": "adaptive" },
80
+ "reasoning_split": true,
81
+ "reasoning_effort": "high",
82
+ "temperature": 1,
83
+ "top_p": 0.95
84
+ }
85
+ },
70
86
  {
71
87
  "id": "MiniMax-M3",
72
88
  "name": "MiniMax M3 (vision)",
@@ -87,11 +103,11 @@
87
103
  }
88
104
  ```
89
105
 
90
- > **Model IDs are case-sensitive.** Use exactly `MiniMax-M3` (not `minimax-m3` or `MINIMAX-M3`).
106
+ > **Model IDs are case-sensitive.** Use exactly `MiniMax-M3.1-Flash` and `MiniMax-M3` (capital `M`s, lowercase `i`). Both entries can live in the same provider group — delete whichever one you don't need. Do **not** invent suffixed ids or duplicate an existing one: VS Code forwards each `id` as the upstream `model` parameter, and the API only knows the two ids above.
91
107
 
92
108
  ### 2. Select your billing mode
93
109
 
94
- MiniMax exposes the **same** `https://api.minimax.io/v1/chat/completions` endpoint and the **same** `MiniMax-M3` model id to both billing modes. The only field that changes is the API key — the rest of `chatLanguageModels.json` stays identical.
110
+ MiniMax exposes the **same** `https://api.minimax.io/v1/chat/completions` endpoint and the **same** model ids (`MiniMax-M3.1-Flash` / `MiniMax-M3`) to both billing modes. The only field that changes is the API key — the rest of `chatLanguageModels.json` stays identical.
95
111
 
96
112
  | Billing mode | Key name in console | Where to get it | How usage is metered |
97
113
  | ------------- | --------------------------------------- | ----------------------------------------------------------------- | ----------------------------------------------- |
@@ -121,64 +137,57 @@ The same `chatLanguageModels.json` block above is reused unchanged. Only the sec
121
137
 
122
138
  When the Token Plan 5-hour or weekly quota is exhausted, the request fails with a Token Plan quota error (see [Troubleshooting](#troubleshooting)). At that point you can either wait for the quota window to reset, buy a Credits top-up at the same console (Credits use the same Subscription Key), or swap the key back to your Open Platform API Key to fall back to PAYG. The URL and `requestBody` do not change during the swap.
123
139
 
124
- ### 4. M3 Priority tier (optional)
125
-
126
- MiniMax exposes M3 with a per-request **`service_tier`** switch that promotes the request into MiniMax's priority admission queue. Priority is **not** a separate model — it is the same `MiniMax-M3` weights invoked with `"service_tier": "priority"` in the request body, costing **1.5× Standard** in exchange for **priority admission** (faster responses, fewer failures during MiniMax peak hours — typically 15:00–17:30 weekdays). Capabilities, context window (1M, guaranteed 512K), vision, tool calling, rate limits (200 RPM / 10M TPM), and thinking modes are identical to Standard.
140
+ ### 4. Thinking depth on M3.1 Flash
127
141
 
128
- **You can only toggle `service_tier` by editing the single `MiniMax-M3` entry.** VS Code's built-in custom-endpoint handler forwards the model block's `id` field as the upstream `model` parameter, and `MiniMax-M3` is the only valid MiniMax M3 model id on the upstream API. Adding a second model block with a different `id` (such as `MiniMax-M3-Priority`) is **not valid** — the upstream API will reject the unknown model. Duplicating the `MiniMax-M3` id twice is also invalid and causes VS Code picker bugs.
129
-
130
- To switch the existing entry from Standard to Priority, add one line to `requestBody`:
142
+ M3.1 Flash always thinks — `thinking: { "type": "disabled" }` and `reasoning_effort: "none"` both return HTTP 400. The only knob is **`reasoning_effort`**, which accepts `low` / `medium` / `high` / `xhigh` / `max` and defaults to `max` when omitted. Lower levels produce fewer thinking tokens at lower latency; `high` is a good starting point for agent/coding work in VS Code, and `max` matches the configuration Artificial Analysis benchmarked.
131
143
 
132
144
  ```json
133
- {
134
- "id": "MiniMax-M3",
135
- "name": "MiniMax M3 (Priority)",
136
- "url": "https://api.minimax.io/v1/chat/completions",
137
- "toolCalling": true,
138
- "vision": true,
139
- "streaming": true,
140
- "maxInputTokens": 1048576,
141
- "maxOutputTokens": 131072,
142
- "requestBody": {
143
- "thinking": { "type": "adaptive" },
144
- "reasoning_split": true,
145
- "temperature": 1,
146
- "top_p": 0.95,
147
- "service_tier": "priority"
148
- }
145
+ "requestBody": {
146
+ "thinking": { "type": "adaptive" },
147
+ "reasoning_split": true,
148
+ "reasoning_effort": "high",
149
+ "temperature": 1,
150
+ "top_p": 0.95
149
151
  }
150
152
  ```
151
153
 
152
- To go back to Standard, remove the `"service_tier": "priority"` line (the API default is `standard`). The `name` field is a local VS Code picker label, so it can be changed freely to reflect the current tier.
153
-
154
- > The `service_tier` parameter is supported on **both** the OpenAI-compatible and Anthropic-compatible endpoints. The third-party `klarkxy/minimax-vscode` extension surfaces the Standard/Priority toggle as a separate picker entry because it is a native `LanguageModelChatProvider` that hard-codes the upstream model id internally and rewrites the picker id — that override mechanism is not available to the built-in custom-endpoint path.
155
-
156
- For the full breakdown (effective pricing, session cost, when Priority is worth the 50% premium), see [docs/research/minimax-m3-priority.md](../research/minimax-m3-priority.md).
154
+ `reasoning_split: true` keeps thinking out of `content` and on `reasoning_content`. Setting it to `false` is not supported on M3.1 Flash, and the server already returns thinking separately by default. `temperature` and `top_p` are accepted but ignored on the M3 family (fixed at `1` and `0.95`); they are kept here only so every provider block in this repo has the same shape.
157
155
 
158
156
  ## Notes
159
157
 
160
- - **`thinking: adaptive` + `reasoning_split: true`** is the recommended pair: the model decides when to reason, and the server returns reasoning in a structured `reasoning_details` field that keeps `content` clean for VS Code. The model still reasons regardless of `thinking.type` — the setting only changes the response field layout.
161
- - **`temperature: 1`, `top_p: 0.95`** are the recommended sampling values for agent/coding work. M3 accepts `temperature` in `[0, 2]` and `top_p` in `[0, 1]`.
162
- - **Vision is image + video on M3 only.**
158
+ - **`thinking: adaptive` + `reasoning_split: true`** is the recommended pair on both models: the server returns thinking on the `reasoning_content` field, which keeps `content` clean for VS Code. On M3.1 Flash the `thinking` field is effectively fixed at `adaptive` (the server rejects `disabled`); on M3, `disabled` is a soft hint — the model may still reason internally and emit `<think>` tags / `reasoning_content`, so the setting changes the response layout more than the model's actual behavior.
159
+ - **`reasoning_effort` only does something on M3.1 Flash.** M3 ignores it. Values are `low` / `medium` / `high` / `xhigh` / `max`, default `max`; `none` is rejected.
160
+ - **`temperature: 1`, `top_p: 0.95`** are accepted but **ignored** on both models (the M3 family is fixed at those values); `top_p` must still be inside `(0, 1]` if you send it.
161
+ - **Interleaved thinking and tool loops:** MiniMax recommends sending the complete assistant message — including `reasoning_content` — back in history so the reasoning chain stays continuous. VS Code does not preserve `reasoning_content` across tool turns, so expect shallower continuity on long agent runs. This is a quality trade-off, not an error: unlike Kimi K2.7, M3.1 Flash does not reject a tool result whose assistant message lacks reasoning content.
162
+ - **Max output:** `max_completion_tokens` is recommended at 131072 and accepts up to 524288 on both models; thinking tokens count toward that limit.
163
163
  - **Rate limits:** 200 RPM / 10M TPM. Input tokens above 512K are available in limited quantity for a limited time.
164
164
 
165
165
  ## Troubleshooting
166
166
 
167
- | Symptom | Likely cause | Fix |
168
- | -------------------------------------------- | ---------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
169
- | Model not in picker | Config not reloaded, or wrong casing | Restart VS Code; verify model ID is exactly `MiniMax-M3` |
170
- | Reasoning leaks into `content` | Missing `reasoning_split` | Add `reasoning_split: true` to `requestBody` |
171
- | 401 Unauthorized | API key region mismatch | Use the endpoint that matches your key's region |
172
- | 401 `token is unusable` / 1004 | Wrong key type on the endpoint — PAYG key against a Token Plan quota or vice versa | The two keys are not interchangeable. Switch the secret in **Chat: Manage Language Models** to the correct key for the billing mode you intend to use. |
173
- | 402 / 429 after a burst of agent turns | Token Plan 5-hour or weekly quota exhausted; not a PAYG balance issue | Wait for the quota window to reset, buy a Credits top-up at `platform.minimax.io/user-center/payment/credits`, upgrade plan, or swap the secret to a PAYG key to continue. |
174
- | 429 rate-limited | Concurrent sessions exceeded 200 RPM / 10M TPM | Reduce concurrent agent sessions |
175
- | Vision request returns 400 | Vision only supported on M3 (not the legacy M2.x line) | Use `MiniMax-M3` |
176
- | Token Plan key works in curl but not VS Code | VS Code is still resolving the old `${input:chat.lm.secret.<id>}` reference | Re-open **Chat: Manage Language Models**, re-paste the Subscription Key, then restart VS Code |
177
- | China account gets 401 on `api.minimax.io` | Key region is China but the URL points to international (or vice versa) | Switch the `url` to `https://api.minimaxi.com/v1/chat/completions` for China, and get the Subscription Key from `platform.minimaxi.com` |
167
+ | Symptom | Likely cause | Fix |
168
+ | -------------------------------------------- | ----------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
169
+ | Model not in picker | Config not reloaded, or wrong casing | Restart VS Code; verify the model ID is exactly `MiniMax-M3.1-Flash` or `MiniMax-M3` |
170
+ | Reasoning leaks into `content` | Missing `reasoning_split` | Add `reasoning_split: true` to `requestBody` |
171
+ | 400 `requires adaptive thinking` (2013) | `thinking: { "type": "disabled" }` or `reasoning_effort: "none"` sent to M3.1 Flash | Remove the override; lower `reasoning_effort` instead. On M3.1 Flash thinking cannot be disabled. |
172
+ | 400 `reasoning_split` cannot be false | `reasoning_split: false` sent to M3.1 Flash | Use `reasoning_split: true` (or omit it); split output is mandatory on this model. |
173
+ | 400 / 2013 with `service_tier` | `service_tier` sent to M3.1 Flash | Remove `service_tier` — only `MiniMax-M3` accepts that field |
174
+ | 401 Unauthorized | API key region mismatch | Use the endpoint that matches your key's region |
175
+ | 401 `token is unusable` / 1004 | Wrong key type on the endpoint — PAYG key against a Token Plan quota or vice versa | The two keys are not interchangeable. Switch the secret in **Chat: Manage Language Models** to the correct key for the billing mode you intend to use. |
176
+ | 402 / 429 after a burst of agent turns | Token Plan 5-hour or weekly quota exhausted; not a PAYG balance issue | Wait for the quota window to reset, buy a Credits top-up at `platform.minimax.io/user-center/payment/credits`, upgrade plan, or swap the secret to a PAYG key to continue. |
177
+ | 429 rate-limited | Concurrent sessions exceeded 200 RPM / 10M TPM | Reduce concurrent agent sessions |
178
+ | Vision request returns 400 | Vision is not supported on the legacy M2.x line | Use `MiniMax-M3.1-Flash` or `MiniMax-M3` |
179
+ | Token Plan key works in curl but not VS Code | VS Code is still resolving the old `${input:chat.lm.secret.<id>}` reference | Re-open **Chat: Manage Language Models**, re-paste the Subscription Key, then restart VS Code |
180
+ | China account gets 401 on `api.minimax.io` | Key region is China but the URL points to international (or vice versa) | Switch the `url` to `https://api.minimaxi.com/v1/chat/completions` for China, and get the Subscription Key from `platform.minimaxi.com` |
178
181
 
179
182
  ## Pricing
180
183
 
181
- For the cross-provider comparison, see [docs/pricing.md](../pricing.md). MiniMax-M3 pay-as-you-go rates (effective — post-50%-off):
184
+ For the cross-provider comparison, see [docs/pricing.md](../pricing.md). MiniMax-M3.1-Flash pay-as-you-go rates are **flat** — one rate for every prompt length, with a 90% cache discount and no `service_tier` support:
185
+
186
+ | Token range | Input (Cache Hit) | Input (Cache Miss) | Output |
187
+ | ------------ | ----------------- | ------------------ | ---------- |
188
+ | All requests | $0.015 / 1M | $0.15 / 1M | $0.60 / 1M |
189
+
190
+ The `MiniMax-M3` rates below are **effective** rates — a permanent 50% off discount halves the list prices:
182
191
 
183
192
  ### Standard tier
184
193
 
@@ -187,18 +196,9 @@ For the cross-provider comparison, see [docs/pricing.md](../pricing.md). MiniMax
187
196
  | ≤ 512K input tokens | $0.06 / 1M | $0.30 / 1M | $1.20 / 1M |
188
197
  | > 512K input tokens\* | $0.12 / 1M | $0.60 / 1M | $2.40 / 1M |
189
198
 
190
- ### Priority tier (`"service_tier": "priority"`)
191
-
192
- Priority is the **same `MiniMax-M3` model** invoked with the `service_tier` admission preference. It costs **1.5× Standard** and earns priority admission during MiniMax peak hours.
193
-
194
- | Token range | Input (Cache Hit) | Input (Cache Miss) | Output |
195
- | --------------------- | ----------------- | ------------------ | ---------- |
196
- | ≤ 512K input tokens | $0.09 / 1M | $0.45 / 1M | $1.80 / 1M |
197
- | > 512K input tokens\* | $0.18 / 1M | $0.90 / 1M | $3.60 / 1M |
198
-
199
199
  \* Input tokens above 512K are available in limited quantity for a limited time.
200
200
 
201
- > **Permanent 50% off:** A standing 50% discount applies to all MiniMax-M3 pay-as-you-go usage on both the Standard and Priority tiers (verified June 9, 2026). The effective rates above are post-discount. List prices (pre-discount) are 2× higher: Standard ≤512K is $0.12 / $0.60 / $2.40; Priority ≤512K is $0.18 / $0.90 / $3.60. See [docs/research/minimax-m3-priority.md](../research/minimax-m3-priority.md) for the full breakdown of when Priority is worth the 50% premium.
201
+ > **Permanent 50% off:** A standing 50% discount applies to all MiniMax-M3 pay-as-you-go usage (verified June 9, 2026). The effective rates above are post-discount. List prices (pre-discount) are 2× higher: $0.12 / $0.60 / $2.40 per 1M at ≤512K input.
202
202
 
203
203
  ### Token Plan (subscription)
204
204
 
@@ -122,14 +122,14 @@ Set the API key through **Chat: Manage Language Models**, then verify the proxy
122
122
  2. **Start the proxy:** `npm run proxy:qwen` (or `npx copilot-custom-endpoint qwen`).
123
123
  3. **Use the proxy-path JSON snippet** below.
124
124
  4. **Set your DashScope API key** via Command Palette → **Chat: Manage Language Models**.
125
- 5. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
125
+ 5. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
126
126
  6. **Restart VS Code.** Reasoning will be visible in plain chat and suppressed on tool turns.
127
127
 
128
128
  ### Direct (no proxy)
129
129
 
130
130
  1. **Use the direct-path JSON snippet** below and replace `<your-workspace-id>` in `url` with your workspace ID.
131
131
  2. **Set your `DASHSCOPE_API_KEY`** via Command Palette → **Chat: Manage Language Models**.
132
- 3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
132
+ 3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4.1 Flash or MiMo V2.6 Flash). [Why?](../../README.md#4-configure-the-utility-small-model)
133
133
  4. **Restart VS Code** and pick "Qwen 3.8 Max".
134
134
 
135
135
  ## Setup
package/docs/pricing.md CHANGED
@@ -1,24 +1,10 @@
1
1
  # Pricing
2
2
 
3
- > **Latest refresh (October 16, 2026):** Muse Spark 1.3 (Meta's closed, API-only `muse-spark-1.3`, released September 2, 2026 on the Meta Model API and Muse Code) is added at **$1.25 / $0.15 cached input / $4.25** per 1M tokens with an Artificial Analysis v4.3.2 composite of **48** (measured at max reasoning effort; the `xhigh` effort variant scores 45), so at **~$0.0219** per intelligence point it sorts between Gemini 3.8 Flash and GLM 5.3. Meta's day-one **Contributor** SKU (`muse-spark-1.3-contributor`, $0.10 / $0.002 / $0.20 per 1M) is excluded from the table because Meta may train future models on those prompts and completions. All other composites and first-party rates were re-checked with no changes.
4
- >
5
- > **Earlier refresh (October 14, 2026):** Gemini 4 Argon (Google's `gemini-4-argon`; Gemini API GA and Copilot-native the same day, October 13, 2026, for Copilot Pro+, Max, Business, and Enterprise) is added at **$2 / $0.10 implicit cache / $10** per 1M tokens — the introductory pricing Google guarantees only through November 30, 2026, after which it doubles to **$4 / $0.20 / $20**. Its Artificial Analysis v4.3.2 composite is **53**, so at **~$0.0377** per intelligence point it sorts between Grok 4.6 and GPT-6.1 Sol. Output pricing includes thinking tokens, and only implicit caching is supported (a flat 95% cache discount). All other composites and first-party rates were re-checked against live provider and OpenRouter metadata with no changes.
6
- >
7
- > **Earlier refresh (October 8, 2026):** Claude Haiku 5.5 (released October 7, 2026; GA in GitHub Copilot the same day for Pro, Pro+, Max, Business, and Enterprise) is added at **$0.10 / $0.01 cache hit / $0.50** per 1M tokens for prompts up to 100,000 tokens, rising to **$0.50 / $0.05 / $2.50** above that threshold, with an Artificial Analysis v4.3.2 composite of **43**. At **~$0.0023** per intelligence point it becomes the lowest cost per intelligence in the table, just ahead of MiMo V2.6 Flash. The same Anthropic announcement halved Claude Sonnet 5.5's cache-read price to **$0.10** per 1M (from $0.20); its Cost per intelligence is unchanged because the modeled session cost does not include cached input. All other composites and first-party rates were re-checked against live OpenRouter metadata and provider pages with no changes.
8
- >
9
- > **Earlier refresh (October 7, 2026):** targeted re-check of the two Qwen entries against OpenRouter's model metadata and ranking card. The rankings card now places unversioned Qwen3.8 Max at **#3** (53.4, tied on score with Claude Fable 5.1), the `qwen/qwen3.8-max-0902` metadata row still reports **45.4**, and OpenRouter lists a newer higher-throughput `qwen/qwen3.8-max-prime` SKU (0923) at **$4 / $0.50 cache read / $12** per 1M — it has no AA composite yet, so it carries no Cost per intelligence row (see [docs/models/qwen.md](models/qwen.md)). No table values changed.
10
- >
11
- > **Earlier refresh (September 30, 2026):** GPT-6.1 Sol (released September 29 at OpenAI DevDay; GA in GitHub Copilot the same day for Pro+, Max, Business, and Enterprise) is added at **$2 / $0.10 cache hit / $10** per 1M tokens — identical input/output to GPT-6 Sol, with cache hits at half the rate ($0.10 vs $0.20) — and an Artificial Analysis v4.3.2 composite of **51.8**, so at **~$0.0386** per intelligence point it sorts between Grok 4.6 and GPT-6 Sol. OpenAI cancelled the planned **GPT-6.1 Astra** over safety concerns, so no such row exists. Every other composite and first-party rate was re-checked against live OpenRouter metadata and provider pages with no changes.
12
- >
13
- > **Earlier refresh (September 29, 2026):** Claude Sonnet 5.5 (released September 28, GA in GitHub Copilot from Pro) is added at **$2 / $0.20 cache hit / $10** per 1M tokens, unchanged from Sonnet 5, with an Artificial Analysis v4.3.2 composite of **56.0**. MiMo V2.6 Flash now has a composite (**37.9**), so it is scored and sorted instead of listed last; at ~$0.0026 it is the lowest cost per intelligence in the table. All other first-party rates were re-verified with no changes.
14
- >
15
- > **Earlier refresh (September 23, 2026):** OpenRouter's current metadata uses Artificial Analysis v4.3.2 composites. The September 22 Copilot releases add Claude Opus 5.5 (**57.6**), GPT-6 Sol (**47.5**), and GPT-6 Luna (**37.3**) to the roster. The API list rates are **$4 / $0.20 cache hit / $20** for Opus 5.5, **$2 / $0.20 / $10** for Sol, and **$0.10 / $0.01 / $0.50** for Luna. MiMo V2.6 Pro is **46.3**, Grok 4.7 is **46.4**, and current values for existing rows are refreshed below. Gemini 3.8 Flash is also available as a GitHub Copilot native model. Xiaomi released the V2.6 Pro, Flash, and Pro UltraSpeed models on September 22; the September 23 roster cleanup removes two retired legacy models. xAI released Grok 4.7 on September 21 at the same short-context rates as Grok 4.6: **$2 / $0.50 cache hit / $6**. The table preserves Qwen's 0803 ranking-card value (**53.4**) and 0902 metadata value (**45.4**) as separate records. The current table is authoritative for present comparisons.
16
-
17
3
  > **⏰ June 1, 2026 — GitHub Copilot switched to usage-based billing (AI Credits) today.**
18
4
  >
19
5
  > Before this change, Copilot used **premium request-based billing** — each model had its own multiplier, and every request consumed `multiplier × 1` from your monthly premium-request allowance. Now **every interaction burns AI credits** based on actual token consumption. Agent mode and complex multi-file tasks consume significantly more tokens than simple Q&A, which means your 7,000 Pro+ credits can disappear fast if you're using frontier models.
20
6
  >
21
- > **The practical workaround:** use cheaper alternative models (DeepSeek V4 Flash, MiMo, Qwen) that are still powerful enough for coding — often at **5–55× less cost** than the Copilot defaults. The tables below show the exact comparison.
7
+ > **The practical workaround:** use cheaper alternative models (DeepSeek V4.1 Flash, MiMo, Qwen) that are still powerful enough for coding — often at **5–55× less cost** than the Copilot defaults. The tables below show the exact comparison.
22
8
  >
23
9
  > 1 AI credit = $0.01 USD. All paid plans include a monthly credit allowance:
24
10
  >
@@ -30,7 +16,7 @@
30
16
  >
31
17
  > Code completions remain unlimited and **not** billed. Auto model selection gets a 10% discount.
32
18
 
33
- All prices below are in **USD per 1M tokens** (non-cached). To convert to AI credits, multiply by 100 (e.g., $5.00/1M = 500 credits/1M). DeepSeek rows use peak rates for the comparison; its official off-peak rates are half. Provider promotions are shown separately in the notes; OpenRouter's effective, batch, or free-provider prices are excluded.
19
+ All prices below are in **USD per 1M tokens** (non-cached). To convert to AI credits, multiply by 100 (e.g., $5.00/1M = 500 credits/1M). The DeepSeek row uses peak rates for the comparison; its official off-peak rates are half. Provider promotions are shown separately in the notes; OpenRouter's effective, batch, or free-provider prices are excluded.
34
20
 
35
21
  All models are listed together below, sorted by Cost per intelligence ascending (lower is better). Models without a Cost per intelligence are ordered by estimated session cost ascending. Session cost assumes ~10K input + ~2K output tokens per turn, 50 turns.
36
22
 
@@ -40,20 +26,20 @@ All models are listed together below, sorted by Cost per intelligence ascending
40
26
  | **MiMo V2.6 Flash** | Xiaomi | **~$0.0026** | **37.9** | $0.14 | $0.0028 | $0.28 | ~$0.10 | 1M |
41
27
  | **GPT-6 Luna** | OpenAI | **~$0.0027** | **37.3** | $0.10 | $0.01 | $0.50 | ~$0.10 | 1M |
42
28
  | **GLM 5.3 Flash** | Z.ai | **~$0.0030** | **41.8** | $0.15 | $0.03 | $0.50 | ~$0.13 | 1M |
29
+ | **MiniMax M3.1 Flash** | MiniMax | **~$0.0031** | **43** | $0.15 | $0.015 | $0.60 | ~$0.14 | 1M |
43
30
  | **GPT-5.6 Luna** | OpenAI | **~$0.0059** | **37.3** | $0.20 | $0.02 | $1.20 | ~$0.22 | 1M |
31
+ | **DeepSeek V4.1 Flash 1017** | DeepSeek | **~$0.0060** | **45** | $0.30 | $0.006 | $1.20 | ~$0.27 | 1M |
44
32
  | **MiMo V2.6 Pro** | Xiaomi | **~$0.0066** | **46.3** | $0.435 | $0.0036 | $0.87 | ~$0.30 | 1M |
45
- | **DeepSeek V4.1 Flash** | DeepSeek | **~$0.0068** | **39.5** | $0.30 | $0.006 | $1.20 | ~$0.27 | 1M |
46
33
  | **MiniMax M3** | MiniMax | **~$0.0093** | **29.2** | $0.60 (≤512K) / $1.20 (>512K) | $0.12 (≤512K) / $0.24 (>512K) | $2.40 (≤512K) / $4.80 (>512K) | ~$0.27 | 1M |
47
- | **MiniMax M3 Priority** | MiniMax | **~$0.0139** | **29.2** | $0.90 (≤512K) / $1.80 (>512K) | $0.18 (≤512K) / $0.36 (>512K) | $3.60 (≤512K) / $7.20 (>512K) | ~$0.41 | 1M |
48
34
  | **Gemini 3.8 Flash** | Google | **~$0.0183** | **40.9** | $0.75 | $0.075 | $3.75 | ~$0.75 | 1M |
49
35
  | **Muse Spark 1.3** | Meta | **~$0.0219** | **48** | $1.25 | $0.15 | $4.25 | ~$1.05 | 1M |
50
36
  | **GLM 5.3** | Z.ai | **~$0.0254** | **44.8** | $1.40 | $0.26 | $4.40 | ~$1.14 | 1M |
51
- | **DeepSeek V4 Pro 0813** | DeepSeek | **~$0.0294** | **36.0** | $1.32 | $0.044 | $3.96 | ~$1.06 | 1M |
52
37
  | **Qwen 3.8 Max (0803)** | DashScope | **~$0.0300** | **53.4** | $2.00 | $0.25 | $6.00 | ~$1.60 | 1M |
53
38
  | **Grok 4.7** | xAI | **~$0.0345** | **46.4** | $2.00 | $0.50 | $6.00 | ~$1.60 | 500K |
54
39
  | **Qwen 3.8 Max (0902)** | DashScope | **~$0.0352** | **45.4** | $2.00 | $0.25 | $6.00 | ~$1.60 | 1M |
55
40
  | **Claude Sonnet 5.5** | Anthropic | **~$0.0357** | **56.0** | $2.00 | $0.10 | $10.00 | ~$2.00 | 1M |
56
41
  | **Grok 4.6** | xAI | **~$0.0361** | **44.3** | $2.00 | $0.50 | $6.00 | ~$1.60 | 500K |
42
+ | **Kimi K3.1 Pro** | Moonshot | **~$0.0363** | **55.1** | $2.00 | $0.20 | $10.00 | ~$2.00 | 1M |
57
43
  | **Gemini 4 Argon** | Google | **~$0.0377** | **53** | $2.00 | $0.10 | $10.00 | ~$2.00 | 1M |
58
44
  | **GPT-6.1 Sol** | OpenAI | **~$0.0386** | **51.8** | $2.00 | $0.10 | $10.00 | ~$2.00 | 1M |
59
45
  | **GPT-6 Sol** | OpenAI | **~$0.0421** | **47.5** | $2.00 | $0.20 | $10.00 | ~$2.00 | 1M |
@@ -72,16 +58,16 @@ Cost per intelligence = estimated session cost ÷ Intelligence Index score. Sess
72
58
 
73
59
  > **Notes:**
74
60
  >
75
- > - **DeepSeek** input pricing shown is the **peak cache miss** price. Peak cache hits are $0.006/M for V4.1 Flash and $0.044/M for Pro; off-peak rates are half. See the [DeepSeek pricing page](https://api-docs.deepseek.com/quick_start/pricing) for the full schedule.
61
+ > - **DeepSeek** input pricing shown is the **peak cache miss** price. Peak cache hits are $0.006/M for V4.1 Flash; off-peak rates are half. The October 17, 2026 **V4.1-Flash-1017** build keeps those rates unchanged, is served under the same `deepseek-flash` ID, and supports both thinking (default) and non-thinking modes. AA measures it at 216 tokens/s and very verbose (210M output tokens to run the Intelligence Index versus a 140M median), so real sessions can cost more than this table's modeled token counts. See the [DeepSeek pricing page](https://api-docs.deepseek.com/quick_start/pricing) for the full schedule.
76
62
  > - **MiMo** input pricing shown is the **cache miss** price. Cache hits are essentially free for V2.6 Pro ($0.0036/M, ~120× cheaper), V2.6 Flash ($0.0028/M, ~50× cheaper), and V2.6 Pro UltraSpeed ($0.036/M, ~120× cheaper). A Xiaomi price cut took effect on 2026-05-27.
77
63
  > - **Anthropic (Claude)** cache reads and cache writes are billed separately. Claude Opus 5.5 cache reads are $0.20/M and Sonnet 5.5 cache reads are $0.10/M (both 0.05× base input); cache writes are $5/M for 5 minutes or $8/M for 1 hour on Opus 5.5 and $2.50/M or $4/M on Sonnet 5.5. Fast mode costs $8/M input and $40/M output. Opus 5.5, Opus 5, Sonnet 5 / 5.5, Haiku 5.5, and Fable 5 / 5.1 use a newer tokenizer that produces approximately 30% more tokens for the same text. **Claude Haiku 5.5 is the only current Claude model with prompt-length tiering**: prompts up to 100,000 tokens cost $0.10 / $0.01 cache read / $0.50 per 1M, prompts over 100,000 tokens cost $0.50 / $0.05 / $2.50, cache writes are $0.125/M (5 min) and $0.20/M (1 h) below the threshold, and batch pricing is $0.05 / $0.25 (≤100K) and $0.25 / $1.25 (>100K); the table above shows the sub-100K rates that Anthropic says cover ~90% of Haiku-class requests. Anthropic states Haiku 5.5 costs about 75% less on average to run than Haiku 4.5. Claude Sonnet 5.5 keeps Sonnet 5's $2 / $10 rates, and Anthropic made them permanent: the Sonnet 5 increase to $3 / $15 scheduled for September 1 was cancelled — and as of October 7, 2026 its cache reads were halved to $0.10/M (batch $1 / $5). Artificial Analysis measured Sonnet 5.5 at 138.7 tokens/s but very verbose (410M output tokens to run the Intelligence Index vs an 88M median), and Haiku 5.5 at 244 tokens/s but also very verbose (440M vs a 100M median), so real sessions can cost more than this table's modeled token counts; Anthropic reports up to 30% fewer tokens per task than Sonnet 5.
78
64
  > - **OpenAI** models support cached input at 0.1× base input rate — except **GPT-6.1 Sol**, whose cache hits are 0.05× ($0.10/M, half of GPT-6 Sol's $0.20/M). GPT-6 Sol, GPT-6.1 Sol, and Luna also bill cache writes separately at 1.25× input. The GPT-6 and GPT-6.1 rows show standard rates through 272K input; for requests >272K, input and cache-read rates double and output is 1.5×.
79
65
  > - **Google (Gemini)** — **Gemini 4 Argon** is priced at Google's introductory rates: $2 / $0.10 implicit cache / $10 per 1M tokens through November 30, 2026, then $4 / $0.20 / $20. Output prices include thinking tokens. Only implicit caching is supported — a flat 95% discount on cached input for prompts of at least 4,096 tokens (explicit caching is not available). Gemini 4 Argon is not offered on the free tier or the Google AI Pro plan, and GitHub Copilot bills it at provider list pricing (Pro+, Max, Business, and Enterprise; not Pro).
80
66
  > - **Qwen** models use **tiered pricing** — determined by total input tokens per request. Prices above are for non-thinking mode.
81
- > - **Kimi** official tables list **Cache Hit before Cache Miss** (opposite order to our table; Kimi K3 also bills cache writes separately at $3.00/M for a 5-minute TTL and $6.00/M for 1 hour). The rows below transpose them so "Input" = cache miss and "Cached input" = cache hit.
67
+ > - **Kimi** official tables list **Cache Hit before Cache Miss** (opposite order to our table; Kimi K3 and K3.1 Pro also bill cache writes separately by TTL — K3 writes are $3.00/M for a 5-minute TTL and $6.00/M for 1 hour). The rows transpose them so "Input" = cache miss and "Cached input" = cache hit. **Kimi K3.1 Pro** (October 23, 2026) is **$2.00 / $0.20 cached / $10.00** per 1M with flat rates and automatic prefix caching (90% cache discount). It is always-thinking: `reasoning_effort` accepts `low` / `high` / `max` (default `max`) and chain-of-thought cannot be disabled. Artificial Analysis measures it at ~81 tokens/s and very verbose (150M output tokens to run the Intelligence Index versus an 87M median), at $2.11 per index task, so real sessions can cost more than the modeled ~$2.00. See [docs/models/kimi.md](models/kimi.md). **Kimi K3** (July 2026) is the previous flagship — superseded by K3.1 Pro but still served at $3.00 / $0.30 cached / $15.00.
82
68
  > - **DashScope** offers a **free quota** of 1M input + 1M output tokens per model, valid for 90 days.
83
- > - **MiniMax M3** uses **tiered pricing** — input price doubles above 512K input tokens. Cache hits are priced at 20% of the input rate ($0.12/M ≤512K, $0.24/M >512K). A **permanent 50% off** discount applies to all MiniMax-M3 pay-as-you-go usage (Standard and Priority tiers), making the effective rates half the list prices above.
84
- > - **MiniMax M3 Priority** is not a separate model — it is the same `MiniMax-M3` weights invoked with `"service_tier": "priority"` in the request body. Priority costs **1.5× Standard** across input, output, and cache reads, in exchange for **priority admission** (faster responses, fewer failures during MiniMax peak hours — typically 15:00–17:30 weekdays). Capabilities, context window (1M, guaranteed 512K), vision, tool calling, rate limits (200 RPM / 10M TPM), and thinking modes are identical to Standard. See [docs/research/minimax-m3-priority.md](research/minimax-m3-priority.md) for the full breakdown.
69
+ > - **MiniMax M3** uses **tiered pricing** — input price doubles above 512K input tokens. Cache hits are priced at 20% of the input rate ($0.12/M ≤512K, $0.24/M >512K). A **permanent 50% off** discount applies to all MiniMax-M3 pay-as-you-go usage, making the effective rates half the list prices above.
70
+ > - **MiniMax M3.1 Flash** is priced **flat** at $0.15 / $0.015 cache read / $0.60 per 1M tokens — one rate for all prompt lengths, a 90% cache discount, and **no `service_tier` support** (the API rejects that field for this model). Thinking cannot be disabled: `thinking: { "type": "disabled" }` or `reasoning_effort: "none"` returns HTTP 400, and depth is tuned with `reasoning_effort` (`low` / `medium` / `high` / `xhigh` / `max`, default `max`). Artificial Analysis measures it at 96 tokens/s and very verbose (180M output tokens to run the Intelligence Index vs a 100M median), so real sessions can cost more than the modeled ~$0.14. See [docs/models/minimax.md](models/minimax.md).
85
71
  > - **GLM** models support prompt caching — cache hits are priced at $0.26/M for 5.3 and $0.03/M for 5.3 Flash (list; $0.015/M during the launch promo).
86
72
  > - **MiMo** offers a **Token Plan** subscription model with discounted rates and a free cache-writing promotion. MiMo V2.6 Flash's **37.9** composite comes from OpenRouter's Artificial Analysis v4.3.2 metadata (AA also measures it at 55.4 tokens/s and very verbose); Xiaomi's published launch benchmarks for it are recorded in [docs/benchmarks.md](benchmarks.md).
87
73
  > - **Meta (Muse Spark)** — the table shows the **Standard** tier of the Meta Model API (`muse-spark-1.3`): $1.25 / $0.15 cached input / $4.25 per 1M tokens, with automatic prefix caching and no long-context surcharge. Meta's day-one **Contributor** SKU (`muse-spark-1.3-contributor`) is $0.10 / $0.002 / $0.20 per 1M — about 12× cheaper on input and 21× on output — but Meta may use those prompts and completions to train future models, so it is excluded from this comparison the way promotions are. Integrated web-search grounding is billed separately at $2.50 per 1,000 queries on both tiers, and rate limits are 3,000 RPM / 4M TPM (Standard) versus 100 RPM / 3M TPM (Contributor). Muse Spark 1.3 is closed and API-only (Meta has announced future open weights without a date), and its audio understanding is degraded versus Muse Spark 1.2 — Meta recommends 1.2 or Muse Voice Transcribe for audio.
@@ -89,7 +75,7 @@ Cost per intelligence = estimated session cost ÷ Intelligence Index score. Sess
89
75
 
90
76
  > **How long does 7,000 credits last?** A Pro+ subscriber running 50-turn sessions can estimate monthly capacity by multiplying each modeled session cost by 100 to convert it to AI credits, then mixing models as needed.
91
77
 
92
- > Prices last verified: October 16, 2026. Always check the official pages for the latest rates:
78
+ > Prices last verified: October 24, 2026. Always check the official pages for the latest rates:
93
79
  >
94
80
  > - [GitHub Copilot models & pricing](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing)
95
81
  > - [OpenAI pricing](https://openai.com/api/pricing/)
@@ -112,7 +98,7 @@ Cost per intelligence = estimated session cost ÷ Intelligence Index score. Sess
112
98
  > - [GitHub Copilot GPT-6.1 Sol release](https://github.blog/changelog/2026-09-29-gpt-6-1-sol-in-github-copilot)
113
99
  > - [GitHub Copilot Claude Opus 5.5 release](https://github.blog/changelog/2026-09-22-claude-opus-5-5-is-now-available-in-github-copilot)
114
100
  > - [GitHub Copilot supported models](https://docs.github.com/en/copilot/reference/ai-models/supported-models)
115
- > - [Artificial Analysis GPT-6 Sol](https://artificialanalysis.ai/models/gpt-6-sol), [GPT-6 Luna](https://artificialanalysis.ai/models/gpt-6-luna), [GPT-6.1 Sol](https://artificialanalysis.ai/models/gpt-6-1-sol), [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5), [Claude Sonnet 5.5](https://artificialanalysis.ai/models/claude-sonnet-5-5), [Claude Haiku 5.5](https://artificialanalysis.ai/models/claude-haiku-5-5), [MiMo-V2.6-Flash](https://artificialanalysis.ai/models/mimo-v2-6-flash), [Gemini 4 Argon](https://artificialanalysis.ai/models/gemini-4-argon), and [Muse Spark 1.3](https://artificialanalysis.ai/models/muse-spark-1-3)
101
+ > - [Artificial Analysis GPT-6 Sol](https://artificialanalysis.ai/models/gpt-6-sol), [GPT-6 Luna](https://artificialanalysis.ai/models/gpt-6-luna), [GPT-6.1 Sol](https://artificialanalysis.ai/models/gpt-6-1-sol), [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5), [Claude Sonnet 5.5](https://artificialanalysis.ai/models/claude-sonnet-5-5), [Claude Haiku 5.5](https://artificialanalysis.ai/models/claude-haiku-5-5), [MiMo-V2.6-Flash](https://artificialanalysis.ai/models/mimo-v2-6-flash), [Gemini 4 Argon](https://artificialanalysis.ai/models/gemini-4-argon), [Muse Spark 1.3](https://artificialanalysis.ai/models/muse-spark-1-3), [Kimi K3.1 Pro](https://artificialanalysis.ai/models/kimi-k3-1-pro), and [MiniMax-M3.1-Flash](https://artificialanalysis.ai/models/minimax-m3-1-flash)
116
102
  > - [Google Gemini pricing](https://ai.google.dev/pricing)
117
103
  > - [Google Gemini 4 Argon (what's new)](https://ai.google.dev/gemini-api/docs/latest-model)
118
104
  > - [Google Gemini 4 Argon model card](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon)
@@ -121,8 +107,11 @@ Cost per intelligence = estimated session cost ÷ Intelligence Index score. Sess
121
107
  > - [DeepSeek pricing](https://api-docs.deepseek.com/quick_start/pricing)
122
108
  > - [MiMo pricing](https://mimo.mi.com/docs/en-US/pricing)
123
109
  > - [MiniMax pricing](https://platform.minimax.io/docs/pricing/overview)
110
+ > - [MiniMax M3.1 Flash announcement](https://www.minimax.io/blog/minimax-m31f)
111
+ > - [MiniMax M3.1 Flash model docs](https://platform.minimax.io/docs/guides/text-generation)
124
112
  > - [Z.ai (GLM) pricing](https://docs.z.ai/guides/overview/pricing)
125
113
  > - [Kimi pricing](https://platform.kimi.ai/docs/pricing)
114
+ > - [Kimi K3.1 Pro quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart)
126
115
  > - [xAI model pricing](https://docs.x.ai/docs/models)
127
116
 
128
117
  - [Meta Model API pricing](https://dev.meta.ai/docs/pricing-rate-limits)
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "copilot-custom-endpoint",
3
- "version": "3.4.4",
3
+ "version": "3.5.0",
4
4
  "description": "Local proxies for VS Code Copilot custom endpoints — Kimi K2, Qwen 3.x & MiMo V2.6",
5
5
  "license": "MIT",
6
6
  "type": "module",
@@ -118,8 +118,9 @@ function rewriteKimi(payload) {
118
118
 
119
119
  // K2.7 and K3 are always-thinking and reject thinking: disabled.
120
120
  // K2.7 uses `thinking` parameter; K3 uses `reasoning_effort` instead.
121
- // Detect K2.7 variants (e.g. kimi-k2.7-code) and K3 variants (e.g. kimi-k3)
122
- // and skip the thinking-disable rewrite while keeping temperature/top_p enforcement.
121
+ // Detect K2.7 variants (e.g. kimi-k2.7-code) and K3-family variants
122
+ // (e.g. kimi-k3, kimi-k3.1-pro, kimi-k3.1-pro-highspeed) and skip the
123
+ // thinking-disable rewrite while keeping temperature/top_p enforcement.
123
124
  const isK27 = model.startsWith('kimi-k2.7')
124
125
  const isK3 = model.startsWith('kimi-k3')
125
126
  const isAlwaysThinking = isK27 || isK3