copilot-custom-endpoint 1.8.1 → 1.9.1

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
@@ -22,6 +22,7 @@ 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** | Moonshot | **Yes** | ✅ | [Manual setup](docs/models/kimi.md) |
25
26
  | **MiMo V2.5** | Xiaomi | Optional (recommended) | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=sdmapvstool.xiaomimimo-for-copilot) ★ / [Manual setup](docs/models/mimo.md) |
26
27
  | **MiMo V2.5 Pro** | Xiaomi | Optional (recommended) | ❌ | [Extension](https://marketplace.visualstudio.com/items?itemName=sdmapvstool.xiaomimimo-for-copilot) ★ / [Manual setup](docs/models/mimo.md) |
27
28
  | **Kimi K2.7 Code / K2.6** | Moonshot | **Yes** | ✅ | [Manual setup](docs/models/kimi.md) |
@@ -117,7 +118,6 @@ All prices are **USD per 1M tokens** (non-cached). 1 AI credit = $0.01. To conve
117
118
  | **MiMo V2.5 Pro** | Xiaomi | **~$0.0072** | **42.2** | ~$0.30 | ❌ | 1M |
118
119
  | **Qwen 3.7 Plus** | DashScope | **~$0.0092** | **39.0** | ~$0.36 | ✅ | 1M |
119
120
  | **MiniMax M3 Priority**⁴ | MiniMax | **~$0.0092** | **44.4** | ~$0.41 | ✅ | 1M |
120
- | **Gemini 3 Flash** | Google | **~$0.020** | **27.0** ³ | ~$0.55 | ✅ | 173K |
121
121
  | **Kimi K2.6** | Moonshot | **~$0.021** | **42.8** | ~$0.88 | ✅ | 262K |
122
122
  | **Kimi K2.7 Code** | Moonshot | **~$0.021** | **42.0** | ~$0.88 | ✅ | 262K |
123
123
  | **GPT-5.4 mini** | OpenAI | **~$0.021** | **40.0** | ~$0.83 | ❌ | 400K |
@@ -125,10 +125,12 @@ All prices are **USD per 1M tokens** (non-cached). 1 AI credit = $0.01. To conve
125
125
  | **GLM 5.2** | Z.ai | **~$0.022** | **51.0** | ~$1.14 | ❌ | 1M |
126
126
  | **GLM 5.1** | Z.ai | **~$0.028** | **40.2** | ~$1.14 | ❌ | 200K |
127
127
  | **GLM 5V Turbo** | Z.ai | **~$0.029** | **34.0** ³ | ~$1.00 | ✅ | 200K |
128
+ | **Gemini 3.6 Flash** ⁸ | Google | **~$0.030** | **50.0** | ~$1.50 | ✅ | 1M |
128
129
  | **Gemini 3.5 Flash** | Google | **~$0.033** | **50.2** | ~$1.65 | ✅ | 1M |
129
130
  | **Qwen 3.7 Max** | DashScope | **~$0.043** | **46.0** | ~$2.00 | ❌ | 1M |
130
131
  | **Gemini 3.1 Pro** | Google | **~$0.047** | **46.5** | ~$2.20 | ✅ | 1M |
131
132
  | **GPT-5.6 Terra** | OpenAI | **~$0.050** | **55.0** | ~$2.75 | ✅ | 1M |
133
+ | **Kimi K3** ⁶ | Moonshot | **~$0.053** | **57.0** | ~$3.00 | ✅ | 1M |
132
134
  | **GPT-5.4** | OpenAI | **~$0.054** | **51.4** | ~$2.75 | ✅ | 1M |
133
135
  | **Claude Sonnet 5** ⁵ | Anthropic | **~$0.057** | **53.0** | ~$3.00 | ✅ | 1M |
134
136
  | **Claude Sonnet 4.6** | Anthropic | **~$0.064** | **47.2** | ~$3.00 | ✅ | 1M |
@@ -153,6 +155,10 @@ All prices are **USD per 1M tokens** (non-cached). 1 AI credit = $0.01. To conve
153
155
 
154
156
  ⁶ **GPT-5.6** launched July 9, 2026. Sol, Terra, and Luna cost $5/$30, $2.50/$15, and $1/$6 per 1M input/output tokens respectively; all support image input and 1M context. Benchmark details and cache-write pricing are maintained in [docs/pricing.md](docs/pricing.md).
155
157
 
158
+ ⁷ **Kimi K3** launched July 16, 2026. 2.8T params (open-source weights by July 27, 2026). Always-thinking reasoning model — uses `reasoning_effort` (not the K2.x `thinking` parameter). AA Intelligence Index score of **57.0** confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/kimi-k3). Priced at $3.00 / $15.00 per MTok input/output. Requires the local Kimi proxy. See [docs/models/kimi.md](docs/models/kimi.md).
159
+
160
+ ⁸ **Gemini 3.6 Flash** launched July 21, 2026. Priced at $1.50 / $7.50 per MTok input/output ($0.15 cached). AA Intelligence Index score of **50.0** confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/gemini-3-6-flash). 1M context with vision.
161
+
156
162
  For footnotes, sources, and detailed notes (cache behavior, tiered pricing, free quotas) see [docs/pricing.md](docs/pricing.md). For a copy-paste config containing **all providers at once**, see [docs/example-config.md](docs/example-config.md).
157
163
 
158
164
  > **👤 Personal picks** —
@@ -164,11 +170,11 @@ For footnotes, sources, and detailed notes (cache behavior, tiered pricing, free
164
170
  > - No local proxy required — works direct
165
171
  > **DeepSeek V4 Pro** is a strong backup — nearly identical intelligence (44.3 vs 44.4) and cost (~$0.0068 vs ~$0.0061 per intelligence point), but noticeably faster in practice.
166
172
  >
167
- > For **plan mode / architecture & design thinking**, **GLM 5.2** is worth the premium:
173
+ > For **plan mode / architecture & design thinking**, **Kimi K3**, **GPT-5.6 Luna**, and **GLM 5.2** are worth the premium:
168
174
  >
169
- > - Highest intelligence score in the table (**51.0**) — best for reasoning-heavy planning
170
- > - 1M lossless context, ideal for whole-codebase planning and large spec review
171
- > - ~$1.14/session is ~4× MiniMax M3, so **not a daily driver for agent mode** — reserve it for planning, then hand implementation off to a cheaper model.
175
+ > - **Kimi K3** — new flagship. Highest intelligence score among custom-endpoint models (**57.0**, #3 overall). Best for complex reasoning, long-horizon planning, and architecture decisions. 1M context with vision. ~$3.00/session is ~11× MiniMax M3, so **not a daily driver** — reserve it for hard problems.
176
+ > - **GPT-5.6 Luna** (Copilot native) — scores **51.2** at ~$1.10/session ($1/$6 per 1M tokens), making it the cheapest model with a 50+ intelligence score. Ideal if you have native Copilot access and want 50+ intelligence on a budget.
177
+ > - **GLM 5.2** — scores **51.0**, reasoning-heavy planning at a lower price. 1M lossless context, ideal for whole-codebase planning and large spec review. ~$1.14/session is ~4× MiniMax M3, so **not a daily driver for agent mode** — reserve it for planning, then hand implementation off to a cheaper model.
172
178
  >
173
179
  > Cheaper alternatives for simpler tasks:
174
180
  >
@@ -40,6 +40,20 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
40
40
  "apiKey": "",
41
41
  "apiType": "chat-completions",
42
42
  "models": [
43
+ {
44
+ "id": "kimi-k3",
45
+ "name": "Kimi K3 (vision)",
46
+ "url": "http://127.0.0.1:3457/v1/chat/completions",
47
+ "requestBody": {
48
+ "temperature": 1,
49
+ "max_tokens": 8192
50
+ },
51
+ "toolCalling": true,
52
+ "vision": true,
53
+ "streaming": true,
54
+ "maxInputTokens": 1000000,
55
+ "maxOutputTokens": 131072
56
+ },
43
57
  {
44
58
  "id": "kimi-k2.6",
45
59
  "name": "Kimi K2.6 (vision)",
@@ -189,7 +203,7 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
189
203
 
190
204
  If you only need one provider, jump straight to its setup guide:
191
205
 
192
- - [Kimi K2.6 / K2.7 Code](kimi.md)
206
+ - [Kimi K3 / K2.6 / K2.7 Code](kimi.md)
193
207
  - [Qwen 3.7 Plus / 3.7 Max](qwen.md)
194
208
  - [Xiaomi MiMo (V2.5 / V2.5 Pro)](mimo.md)
195
209
  - [MiniMax M3 (Standard)](minimax.md) and [MiniMax M3 (Priority)](minimax.md#5-m3-priority-tier-optional)
@@ -1,6 +1,8 @@
1
1
  # Kimi (Moonshot) — VS Code Custom Endpoint Setup Guide
2
2
 
3
- > **TL;DR:** Kimi models require the local proxy. The K2 family locks `temperature: 1` and `top_p: 0.95`. K2.6 needs `thinking: { type: "disabled" }` on tool turns; **K2.7 Code is always-thinking and rejects `thinking: disabled`**, so the proxy detects `kimi-k2.7*` and skips that rewrite while keeping sampling enforcement. Direct VS Code → Moonshot is not viable.
3
+ > **TL;DR:** Kimi models require the local proxy. The K2/K3 family locks `temperature: 1` and `top_p: 0.95`. K2.6 needs `thinking: { type: "disabled" }` on tool turns; **K2.7 Code is always-thinking and rejects `thinking: disabled`**. **K3 is always-thinking and uses `reasoning_effort` (not the `thinking` parameter)** — the proxy detects `kimi-k3` and skips the thinking-disable rewrite while keeping sampling enforcement. Direct VS Code → Moonshot is not viable.
4
+ >
5
+ > **🆕 Kimi K3** (released July 16, 2026) is Kimi's flagship — 2.8T parameters, 1M context, AA Intelligence Index **57.0** (#3 overall, between GPT-5.6 Sol and Claude Opus 4.8). Open-source weights coming by July 27, 2026.
4
6
  >
5
7
  > **🆕 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).
6
8
 
@@ -20,12 +22,13 @@
20
22
 
21
23
  ### Models
22
24
 
23
- | Model | Vision | Context | Max output | Notes |
24
- | ---------------- | ------ | ------- | ---------- | --------------------------------------------------------------------------------- |
25
- | `kimi-k2.6` | ✅ Yes | 262K | 32768 | Proxy forces `thinking: { type: "disabled" }` on tool turns |
26
- | `kimi-k2.7-code` | ✅ Yes | 262K | 4096 | Always-thinking; rejects `thinking: disabled`. Keep `maxOutputTokens` low (4096). |
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 **57.0**. |
28
+ | `kimi-k2.6` | ✅ Yes | 262K | 32768 | Proxy forces `thinking: { type: "disabled" }` on tool turns |
29
+ | `kimi-k2.7-code` | ✅ Yes | 262K | 4096 | Always-thinking; rejects `thinking: disabled`. Keep `maxOutputTokens` low (4096). |
27
30
 
28
- > Deprecated K2 ids (`kimi-k2-0905-preview`, `kimi-k2-turbo-preview`, `kimi-k2-thinking`, etc.) were discontinued May 25, 2026 — use K2.6 or K2.7.
31
+ > Deprecated K2 ids (`kimi-k2-0905-preview`, `kimi-k2-turbo-preview`, `kimi-k2-thinking`, etc.) were discontinued May 25, 2026 — use K2.6, K2.7, or K3.
29
32
 
30
33
  ## Quick Start
31
34
 
@@ -33,7 +36,7 @@
33
36
  2. **Edit `chatLanguageModels.json`** — add the Kimi block below.
34
37
  3. **Set your Moonshot API key** via Command Palette → **Chat: Manage Language Models**.
35
38
  4. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.5). [Why?](../../README.md#4-configure-the-utility-small-model)
36
- 5. **Restart VS Code** and pick "Kimi K2.6" or "Kimi K2.7 Code".
39
+ 5. **Restart VS Code** and pick "Kimi K3", "Kimi K2.6", or "Kimi K2.7 Code".
37
40
 
38
41
  ## Setup
39
42
 
@@ -52,6 +55,17 @@
52
55
  "apiKey": "",
53
56
  "apiType": "chat-completions",
54
57
  "models": [
58
+ {
59
+ "id": "kimi-k3",
60
+ "name": "Kimi K3 (vision)",
61
+ "url": "http://127.0.0.1:3457/v1/chat/completions",
62
+ "requestBody": { "temperature": 1, "max_tokens": 8192 },
63
+ "toolCalling": true,
64
+ "vision": true,
65
+ "streaming": true,
66
+ "maxInputTokens": 1000000,
67
+ "maxOutputTokens": 131072
68
+ },
55
69
  {
56
70
  "id": "kimi-k2.6",
57
71
  "name": "Kimi K2.6 (vision)",
@@ -78,6 +92,8 @@
78
92
  }
79
93
  ```
80
94
 
95
+ > **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.
96
+ >
81
97
  > **K2.7 note:** `max_tokens` and `maxOutputTokens` are intentionally conservative at **4096**. K2.7 is always-thinking, so reasoning tokens inflate response size. Values above 24K triggered VS Code's "Response too long" error in agent mode.
82
98
 
83
99
  ### 2. API key
@@ -93,7 +109,7 @@ Kimi is **Pay-as-You-Go only** — no subscription. A single Open Platform API k
93
109
 
94
110
  ## Local Proxy
95
111
 
96
- The `proxy/kimi-proxy.mjs` is **required** for Kimi — K2-family sampling constraints and tool-turn thinking behaviour cannot be satisfied by VS Code's direct path. See [Quick Start](#quick-start) for the launch command.
112
+ The `proxy/kimi-proxy.mjs` is **required** for Kimi — K2/K3-family sampling constraints and tool-turn thinking behaviour cannot be satisfied by VS Code's direct path. See [Quick Start](#quick-start) for the launch command.
97
113
 
98
114
  | Setting | Value |
99
115
  | ------------ | ----------------------------------------------------- |
@@ -123,33 +139,38 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
123
139
  - Rewrites plain-chat requests to `temperature: 1`, `top_p: 0.95`.
124
140
  - For **K2.5 / K2.6**: rewrites tool-enabled requests to `thinking: { type: "disabled" }`, `temperature: 0.6`, `top_p: 0.95`.
125
141
  - For **K2.7 Code**: keeps thinking enabled (K2.7 rejects `thinking: disabled` with HTTP 400); rewrites to `temperature: 1`, `top_p: 0.95`.
142
+ - 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.
126
143
  - Preserves streaming responses (SSE).
127
144
  - Writes redacted request summaries to `debug_log/kimi-proxy.ndjson`.
128
145
 
129
146
  ## Notes
130
147
 
148
+ - **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).
131
149
  - **K2.7 Code** is always-thinking. `max_tokens` and `maxOutputTokens` are intentionally conservative at **4096** — reasoning tokens inflate response size and values above 24K triggered VS Code's "Response too long" error in agent mode.
132
150
  - **K2 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.
133
- - **`tool_choice` only supports `auto`** — don't override it (VS Code's default is `auto`).
151
+ - **`tool_choice` for K2.x only supports `auto`** — don't override it (VS Code's default is `auto`). K3 also supports `required` and named tools.
152
+ - **K3 `reasoning_effort` vs K2.x `thinking`:** K3 uses `reasoning_effort` (currently only `max` level) at the top level of the request body. Do not mix this with the K2.x `thinking` parameter — K3 rejects it. The proxy preserves whatever `reasoning_effort` VS Code sends and only enforces temperature/top_p.
134
153
 
135
154
  ## Troubleshooting
136
155
 
137
- | Symptom | Likely cause | Fix |
138
- | ------------------------------------------------------ | -------------------------- | ------------------------------------------------- |
139
- | "Connection refused" on chat | Proxy not running | `npm run proxy:kimi` |
140
- | `invalid temperature: only 1 is allowed` | Direct path without proxy | Use the proxy |
141
- | `invalid top_p: only 0.95 is allowed` | Direct path without proxy | Use the proxy |
142
- | `thinking is enabled but reasoning_content is missing` | Tool turn with thinking on | Verify `KIMI_PROXY_DISABLE_THINKING_WITH_TOOLS=1` |
143
- | Model not in VS Code picker | Config not reloaded | Restart VS Code |
144
- | `tool_choice=required` rejected | Model limitation | Use `auto` only |
156
+ | Symptom | Likely cause | Fix |
157
+ | ------------------------------------------------------ | ---------------------------- | ------------------------------------------------- |
158
+ | "Connection refused" on chat | Proxy not running | `npm run proxy:kimi` |
159
+ | `invalid temperature: only 1 is allowed` | Direct path without proxy | Use the proxy |
160
+ | `invalid top_p: only 0.95 is allowed` | Direct path without proxy | Use the proxy |
161
+ | `thinking is enabled but reasoning_content is missing` | Tool turn with thinking on | Verify `KIMI_PROXY_DISABLE_THINKING_WITH_TOOLS=1` |
162
+ | `unsupported parameter: thinking` with K3 | K3 doesn't accept `thinking` | Use `reasoning_effort` instead; proxy detects K3 |
163
+ | Model not in VS Code picker | Config not reloaded | Restart VS Code |
164
+ | `tool_choice=required` rejected (K2.x) | Model limitation | Use `auto` only for K2.x; K3 supports `required` |
145
165
 
146
166
  ## Pricing
147
167
 
148
168
  For the cross-provider comparison, see [docs/pricing.md](../pricing.md). Moonshot direct platform rates:
149
169
 
150
- | Model | Input (cache miss) | Cached input | Output |
151
- | ---------------- | ------------------ | ------------ | ---------- |
152
- | `kimi-k2.6` | $0.95 / 1M | $0.16 / 1M | $4.00 / 1M |
153
- | `kimi-k2.7-code` | $0.95 / 1M | $0.19 / 1M | $4.00 / 1M |
170
+ | Model | Input (cache miss) | Cached input | Output |
171
+ | ---------------- | ------------------ | ------------ | ----------- |
172
+ | `kimi-k3` | $3.00 / 1M | $0.30 / 1M | $15.00 / 1M |
173
+ | `kimi-k2.6` | $0.95 / 1M | $0.16 / 1M | $4.00 / 1M |
174
+ | `kimi-k2.7-code` | $0.95 / 1M | $0.19 / 1M | $4.00 / 1M |
154
175
 
155
- > K2.7 has no non-thinking mode. Via DashScope, K2.6 is also available at $0.89 / 1M input and $3.71 / 1M output.
176
+ > K3 pricing is flat (no tiering by context length). K2.7 has no non-thinking mode. Via DashScope, K2.6 is also available at $0.89 / 1M input and $3.71 / 1M output.
package/docs/pricing.md CHANGED
@@ -1,6 +1,6 @@
1
1
  # Pricing
2
2
 
3
- > **Updated:** July 12, 2026 — GPT-5.6 Sol, Terra, and Luna pricing and benchmark scores added from OpenAI's July 9 release announcement and Artificial Analysis model pages.
3
+ > **Updated:** July 22, 2026 — Gemini 3.6 Flash added (released July 21, 2026). AA Intelligence Index **50.0**, pricing $1.50 / $7.50 per MTok input/output.
4
4
 
5
5
  > **⏰ June 1, 2026 — GitHub Copilot switched to usage-based billing (AI Credits) today.**
6
6
  >
@@ -31,7 +31,6 @@ All models are listed together below, sorted by Cost per intelligence ascending
31
31
  | **MiMo V2.5 Pro** | Xiaomi | **~$0.0072** | **42.2** | $0.435 | $0.0036 | $0.87 | ~$0.30 | 1M |
32
32
  | **Qwen 3.7 Plus** | DashScope | **~$0.0092** | **39.0** | $0.40 (≤256K) / $1.20 (>256K) | — | $1.60 (≤256K) / $4.80 (>256K) | ~$0.36 | 1M |
33
33
  | **MiniMax M3 Priority**⁵ | MiniMax | **~$0.0092** | **44.4** | $0.90 (≤512K) / $1.80 (>512K) | $0.18 (≤512K) / $0.36 (>512K) | $3.60 (≤512K) / $7.20 (>512K) | ~$0.41 | 1M |
34
- | **Gemini 3 Flash** | Google | **~$0.020** | **27.0** ³ | $0.50 | $0.05 | $3.00 | ~$0.55 | 173K |
35
34
  | **Kimi K2.6** | Moonshot | **~$0.021** | **42.8** | $0.95 | $0.16 | $4.00 | ~$0.88 | 256K |
36
35
  | **Kimi K2.7 Code** | Moonshot | **~$0.021** | **42.0** | $0.95 | $0.19 | $4.00 | ~$0.88 | 262K |
37
36
  | **GPT-5.4 mini** | OpenAI | **~$0.021** | **40.0** | $0.75 | $0.075 | $4.50 | ~$0.83 | 400K |
@@ -39,11 +38,13 @@ All models are listed together below, sorted by Cost per intelligence ascending
39
38
  | **GLM 5.2** ⁴ | Z.ai | **~$0.022** | **51.0** | $1.40 | $0.26 | $4.40 | ~$1.14 | 1M |
40
39
  | **GLM 5.1** | Z.ai | **~$0.028** | **40.2** | $1.40 | $0.26 | $4.40 | ~$1.14 | 200K |
41
40
  | **GLM 5V Turbo** | Z.ai | **~$0.029** | **34.0** ³ | $1.20 | $0.24 | $4.00 | ~$1.00 | 200K |
41
+ | **Gemini 3.6 Flash** ⁹ | Google | **~$0.030** | **50.0** | $1.50 | $0.15 | $7.50 | ~$1.50 | 1M |
42
42
  | **Gemini 3.5 Flash** | Google | **~$0.033** | **50.2** | $1.50 | $0.15 | $9.00 | ~$1.65 | 1M |
43
43
  | **Qwen 3.7 Max** | DashScope | **~$0.043** | **46.0** | $2.50 (≤1M) | — | $7.50 (≤1M) | ~$2.00 | 1M |
44
44
  | **Gemini 3.1 Pro** | Google | **~$0.047** | **46.5** | $2.00¹ | $0.20 | $12.00¹ | ~$2.20 | 1M |
45
- | **GPT-5.3-Codex** | OpenAI | **~$0.052** | **44.0** ³ | $1.75 | $0.175 | $14.00 | ~$2.28 | 400K |
46
45
  | **GPT-5.6 Terra** | OpenAI | **~$0.050** | **55.0** | $2.50 | $0.25 | $15.00 | ~$2.75 | 1M |
46
+ | **GPT-5.3-Codex** | OpenAI | **~$0.052** | **44.0** ³ | $1.75 | $0.175 | $14.00 | ~$2.28 | 400K |
47
+ | **Kimi K3** ⁷ | Moonshot | **~$0.053** | **57.0** | $3.00 | $0.30 | $15.00 | ~$3.00 | 1M |
47
48
  | **GPT-5.4** | OpenAI | **~$0.054** | **51.4** | $2.50 | $0.25 | $15.00 | ~$2.75 | 1M |
48
49
  | **Claude Sonnet 5** ⁶ | Anthropic | **~$0.057** | **53.0** | $3.00 | $0.30 | $15.00 | ~$3.00 | 1M |
49
50
  | **Claude Sonnet 4.6** | Anthropic | **~$0.064** | **47.2** | $3.00 | $0.30 | $15.00 | ~$3.00 | 1M |
@@ -67,7 +68,11 @@ All models are listed together below, sorted by Cost per intelligence ascending
67
68
 
68
69
  ⁶ **Claude Sonnet 5** has introductory pricing of **$2.00 / $0.20 / $10.00** (input / cached / output) through **August 31, 2026**, after which the standard pricing of $3.00 / $0.30 / $15.00 shown above takes effect. See [Anthropic's pricing page](https://claude.com/pricing) for the latest.
69
70
 
70
- ⁷ **GPT-5.6** pricing is **$1.00 / $0.10 / $6.00** for Luna, **$2.50 / $0.25 / $15.00** for Terra, and **$5.00 / $0.50 / $30.00** for Sol (input / cached input / output per 1M tokens). OpenAI bills cache writes at 1.25x the uncached input rate for GPT-5.6 and later. All three tiers have a 1M-token context window and support text + image input. See [OpenAI's GPT-5.6 announcement](https://openai.com/index/gpt-5-6/) and the [Artificial Analysis model pages](https://artificialanalysis.ai/models/gpt-5-6-sol).
71
+ ⁷ **Kimi K3** was released on **July 16, 2026**. 2.8T parameters (open-source, weights by July 27, 2026). 1M context window. Always-thinking reasoning model — uses `reasoning_effort` (not the K2.x `thinking` parameter). Fixed sampling: `temperature=1`, `top_p=0.95`. Pricing is flat (no tiering by context length). See [Kimi K3 pricing](https://platform.kimi.ai/docs/pricing/chat-k3) and the [Artificial Analysis model page](https://artificialanalysis.ai/models/kimi-k3).
72
+
73
+ ⁸ **GPT-5.6** pricing is **$1.00 / $0.10 / $6.00** for Luna, **$2.50 / $0.25 / $15.00** for Terra, and **$5.00 / $0.50 / $30.00** for Sol (input / cached input / output per 1M tokens). OpenAI bills cache writes at 1.25x the uncached input rate for GPT-5.6 and later. All three tiers have a 1M-token context window and support text + image input. See [OpenAI's GPT-5.6 announcement](https://openai.com/index/gpt-5-6/) and the [Artificial Analysis model pages](https://artificialanalysis.ai/models/gpt-5-6-sol).
74
+
75
+ ⁹ **Gemini 3.6 Flash** was released on **July 21, 2026**. AA Intelligence Index score (**50.0**) confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/gemini-3-6-flash). Pricing is $1.50 input / $0.15 cached / $7.50 output per MTok.
71
76
 
72
77
  Cost per intelligence = estimated session cost ÷ Intelligence Index score. Session cost assumes ~10K input + ~2K output tokens per turn, 50 turns.
73
78
 
@@ -75,7 +80,6 @@ Cost per intelligence = estimated session cost ÷ Intelligence Index score. Sess
75
80
  >
76
81
  > - **DeepSeek V4** input pricing shown is the **cache miss** price. Cache hits are significantly cheaper ($0.0028/M for Flash, $0.003625/M for Pro).
77
82
  > - **MiMo** input pricing shown is the **cache miss** price. Cache hits are essentially free for V2.5 Pro ($0.0036/M, ~120× cheaper) and V2.5 ($0.0028/M, ~50× cheaper). A Xiaomi price cut took effect on 2026-05-27.
78
- > - **Gemini 3 Flash** is priced at $0.50/MTok input (text/image/video) and $1.00/MTok input for audio.
79
83
  > - **Anthropic (Claude)** models also have a cache write cost ($6.25/MTok for Opus, $3.75/MTok for Sonnet, $1.25/MTok for Haiku). Opus 4.7+, Sonnet 5, and Fable 5 use a new tokenizer that produces approximately 30% more tokens for the same text.
80
84
  > - **OpenAI** models support cached input at 0.1× base input rate.
81
85
  > - **Qwen** models use **tiered pricing** — determined by total input tokens per request. Prices above are for non-thinking mode.
@@ -91,7 +95,7 @@ Cost per intelligence = estimated session cost ÷ Intelligence Index score. Sess
91
95
 
92
96
  > **How long does 7,000 credits last?** A Pro+ subscriber running 50-turn sessions could afford roughly **13 GPT-5.5 sessions**, **23 Opus sessions**, or **212 Raptor mini sessions** per month — or mix and match. (Multiply session cost by 100 to convert to AI credits.)
93
97
 
94
- > Prices last verified: July 12, 2026. Always check the official pages for the latest rates:
98
+ > Prices last verified: July 22, 2026. Always check the official pages for the latest rates:
95
99
  >
96
100
  > - [GitHub Copilot models & pricing](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing)
97
101
  > - [Microsoft MAI-Code-1-Flash model card](https://docs.github.com/en/copilot/reference/ai-models/model-comparison#task-general-purpose-coding-and-writing)
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "copilot-custom-endpoint",
3
- "version": "1.8.1",
3
+ "version": "1.9.1",
4
4
  "description": "Local proxies for VS Code Copilot custom endpoints — Kimi K2, Qwen 3.x & MiMo V2.5",
5
5
  "license": "MIT",
6
6
  "type": "module",
@@ -57,4 +57,4 @@
57
57
  "dependencies": {
58
58
  "dotenv": "^17.4.2"
59
59
  }
60
- }
60
+ }
@@ -12,10 +12,15 @@ import { createProxy } from '../lib/create-proxy.mjs'
12
12
  * always-thinking and rejects `thinking: { type: 'disabled' }`. The proxy
13
13
  * detects K2.7 and skips the thinking-disable rewrite while keeping
14
14
  * temperature/top_p enforcement.
15
+ * - Validated in this repo with `kimi-k3` (July 17, 2026). K3 is always-thinking
16
+ * and uses `reasoning_effort` (NOT the K2.x `thinking` parameter). The proxy
17
+ * detects K3 and skips the thinking-disable rewrite while keeping
18
+ * temperature/top_p enforcement. It does NOT inject a `thinking` block —
19
+ * K3 rejects it.
15
20
  * - Not intended for `moonshot-v1` models or non-Kimi providers, because this
16
- * proxy rewrites requests to K2-family-specific values:
21
+ * proxy rewrites requests to K2/K3-family-specific values:
17
22
  * - thinking mode temperature = 1.0
18
- * - non-thinking mode temperature = 0.6
23
+ * - non-thinking mode temperature = 0.6 (K2.5/K2.6 only)
19
24
  * - top_p = 0.95
20
25
  * - tool-enabled requests force `thinking: { type: 'disabled' }` (K2.5/K2.6 only)
21
26
  */
@@ -108,12 +113,16 @@ function rewriteKimi(payload) {
108
113
  const incomingTemperature = payload.temperature
109
114
  const incomingTopP = payload.top_p
110
115
  const incomingThinkingType = payload?.thinking?.type
116
+ const incomingReasoningEffort = payload?.reasoning_effort
111
117
  const model = payload.model ?? ''
112
118
 
113
- // K2.7 is always-thinking and rejects thinking: disabled.
114
- // Detect K2.7 variants (e.g. kimi-k2.7-code) and skip the thinking-disable
115
- // rewrite while keeping temperature/top_p enforcement.
119
+ // K2.7 and K3 are always-thinking and reject thinking: disabled.
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.
116
123
  const isK27 = model.startsWith('kimi-k2.7')
124
+ const isK3 = model.startsWith('kimi-k3')
125
+ const isAlwaysThinking = isK27 || isK3
117
126
 
118
127
  // Determine if a tool is actually being invoked:
119
128
  // - tool_choice is set and not "none"
@@ -126,13 +135,15 @@ function rewriteKimi(payload) {
126
135
  (toolChoice !== undefined && toolChoice !== 'none' && toolChoice !== null)
127
136
  const hasTools = hasActiveToolCall
128
137
 
129
- const useNonThinkingMode = !isK27 && disableThinkingWithTools && hasTools
138
+ const useNonThinkingMode =
139
+ !isAlwaysThinking && disableThinkingWithTools && hasTools
130
140
  const rewrittenTemperature = useNonThinkingMode
131
141
  ? forcedNonThinkingTemperature
132
142
  : forcedTemperature
133
143
 
134
144
  // Capture incoming state before mutation
135
145
  payload.__incomingThinkingType = incomingThinkingType
146
+ payload.__incomingReasoningEffort = incomingReasoningEffort
136
147
 
137
148
  // Apply rewrites
138
149
  payload.temperature = rewrittenTemperature
@@ -142,26 +153,35 @@ function rewriteKimi(payload) {
142
153
  payload.thinking = { type: 'disabled' }
143
154
  }
144
155
 
156
+ // K3 does not accept `thinking` — delete it if present so it doesn't cause a 400
157
+ if (isK3 && payload.thinking) {
158
+ delete payload.thinking
159
+ }
160
+
145
161
  const rewrittenThinkingType = payload.thinking?.type
146
162
  const rewriteInfo = {
147
163
  model,
148
164
  isK27,
165
+ isK3,
149
166
  incomingTemperature,
150
167
  rewrittenTemperature,
151
168
  incomingTopP,
152
169
  rewrittenTopP: forcedTopP,
153
170
  incomingThinkingType,
154
- rewrittenThinkingType
171
+ rewrittenThinkingType,
172
+ incomingReasoningEffort
155
173
  }
156
174
 
157
175
  const summary = summarizePayload(payload, hasTools, rewriteInfo)
158
176
 
159
177
  const modeTag = hasTools ? '[tools]' : '[chat]'
160
178
  const k27Tag = isK27 ? '[k2.7]' : ''
161
- const consoleMsg = `${k27Tag}${modeTag} temperature ${String(incomingTemperature)} -> ${String(rewrittenTemperature)}, top_p ${String(incomingTopP)} -> ${String(forcedTopP)}, thinking ${String(incomingThinkingType)} -> ${String(rewrittenThinkingType)}`
179
+ const k3Tag = isK3 ? '[k3]' : ''
180
+ const consoleMsg = `${k27Tag}${k3Tag}${modeTag} temperature ${String(incomingTemperature)} -> ${String(rewrittenTemperature)}, top_p ${String(incomingTopP)} -> ${String(forcedTopP)}, thinking ${String(incomingThinkingType)} -> ${String(rewrittenThinkingType)}, reasoning_effort ${String(incomingReasoningEffort)}`
162
181
 
163
- // Clean up internal key before forwarding
182
+ // Clean up internal keys before forwarding
164
183
  delete payload.__incomingThinkingType
184
+ delete payload.__incomingReasoningEffort
165
185
 
166
186
  return { summary, consoleMsg }
167
187
  }