copilot-custom-endpoint 2.2.0 → 3.1.0
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 +42 -114
- package/docs/example-config.md +3 -95
- package/docs/models/glm.md +3 -49
- package/docs/models/kimi.md +13 -43
- package/docs/models/qwen.md +75 -51
- package/docs/pricing.md +40 -76
- package/package.json +1 -1
- package/proxy/qwen-proxy.mjs +9 -2
package/README.md
CHANGED
|
@@ -20,20 +20,18 @@ That's it. No code, no servers to manage (unless the model specifically needs th
|
|
|
20
20
|
|
|
21
21
|
## Pick a model
|
|
22
22
|
|
|
23
|
-
| Model
|
|
24
|
-
|
|
|
25
|
-
| **Kimi K3
|
|
26
|
-
| **MiMo V2.5**
|
|
27
|
-
| **MiMo V2.5 Pro**
|
|
28
|
-
| **Qwen 3.
|
|
29
|
-
| **Qwen 3.
|
|
30
|
-
| **
|
|
31
|
-
| **
|
|
32
|
-
| **GLM 5.3
|
|
33
|
-
| **
|
|
34
|
-
| **
|
|
35
|
-
| **DeepSeek V4 Flash 0731** | 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 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) |
|
|
23
|
+
| Model | Provider | Needs proxy? | Vision | Setup guide |
|
|
24
|
+
| -------------------------- | ---------- | ----------------------- | ------ | ------------------------------------------------------------------------------------------------------------------------------------------- |
|
|
25
|
+
| **Kimi K3** | Moonshot | **Yes** | ✅ | [Manual setup](docs/models/kimi.md) |
|
|
26
|
+
| **MiMo V2.5** | Xiaomi | Optional (recommended) | ✅ | [Extension](https://marketplace.visualstudio.com/items?itemName=sdmapvstool.xiaomimimo-for-copilot) ★ / [Manual setup](docs/models/mimo.md) |
|
|
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) |
|
|
28
|
+
| **Qwen 3.8 Max (0803)** | OpenRouter | Optional (experimental) | ✅ | [Snapshot/proxy setup](docs/models/qwen.md#optional-openrouter-snapshot-proxy) |
|
|
29
|
+
| **Qwen 3.8 Max (0902)** | DashScope | Optional (recommended) | ✅ | [Manual setup](docs/models/qwen.md) |
|
|
30
|
+
| **MiniMax M3** | MiniMax | No | ✅ | [Extension](https://github.com/tugudush/minimax-copilot) ★ / [Manual setup](docs/models/minimax.md) |
|
|
31
|
+
| **GLM 5.3 Flash** | Z.ai | No | ✅ | [Manual setup](docs/models/glm.md) |
|
|
32
|
+
| **GLM 5.3** | Z.ai | No | ❌ | [Manual setup](docs/models/glm.md) |
|
|
33
|
+
| **DeepSeek V4 Flash 0731** | DeepSeek | No (uses an extension) | ❌ | [Extension](https://marketplace.visualstudio.com/items?itemName=Vizards.deepseek-v4-for-copilot) / [setup notes](docs/models/deepseek.md) |
|
|
34
|
+
| **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) |
|
|
37
35
|
|
|
38
36
|
## Setup
|
|
39
37
|
|
|
@@ -107,108 +105,38 @@ npx copilot-custom-endpoint clean # Remove debug_log/
|
|
|
107
105
|
|
|
108
106
|
## Pricing snapshot
|
|
109
107
|
|
|
110
|
-
All prices are **USD per 1M tokens** (non-cached). 1 AI credit = $0.01. To convert to AI credits, multiply by 100 (e.g., $5.00/1M = 500 credits/1M). Session cost assumes ~10K input + ~2K output tokens per turn, 50 turns. DeepSeek V4 estimates use peak rates; official off-peak rates are half. The current AA values come from OpenRouter's
|
|
111
|
-
|
|
112
|
-
| Model
|
|
113
|
-
|
|
|
114
|
-
| **GLM 5.3 Flash**
|
|
115
|
-
| **
|
|
116
|
-
| **
|
|
117
|
-
| **
|
|
118
|
-
| **
|
|
119
|
-
| **
|
|
120
|
-
| **
|
|
121
|
-
| **Gemini 3.
|
|
122
|
-
| **
|
|
123
|
-
| **
|
|
124
|
-
| **
|
|
125
|
-
| **Grok 4.6**
|
|
126
|
-
| **Qwen 3.8 Max (0902)**
|
|
127
|
-
| **
|
|
128
|
-
| **Claude Sonnet 5**
|
|
129
|
-
| **
|
|
130
|
-
| **
|
|
131
|
-
| **
|
|
132
|
-
| **
|
|
133
|
-
| **
|
|
134
|
-
| **Claude Fable 5
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
| **MiMo V2.5** | Xiaomi | — | — | ~$0.10 | ✅ | 1M |
|
|
138
|
-
| **Qwen 3.7 Plus** | DashScope | — | — | ~$0.36 | ✅ | 1M |
|
|
139
|
-
| **Qwen 3.6 Plus** | DashScope | — | — | ~$0.55 | ✅ | 1M |
|
|
140
|
-
| **GPT-5.4 mini** | OpenAI | — | — | ~$0.83 | ❌ | 400K |
|
|
141
|
-
| **Kimi K2.6** | Moonshot | — | — | ~$0.88 | ✅ | 256K |
|
|
142
|
-
| **Kimi K2.7 Code** | Moonshot | — | — | ~$0.88 | ✅ | 262K |
|
|
143
|
-
| **GLM 5V Turbo** | Z.ai | — | — | ~$1.00 | ✅ | 200K |
|
|
144
|
-
| **GLM 5.1** | Z.ai | — | — | ~$1.14 | ❌ | 200K |
|
|
145
|
-
| **GLM 5.2** | Z.ai | — | — | ~$1.14 | ❌ | 1M |
|
|
146
|
-
| **Gemini 3.5 Flash** | Google | — | — | ~$1.65 | ✅ | 1M |
|
|
147
|
-
| **Qwen 3.7 Max** | DashScope | — | — | ~$2.00 | ❌ | 1M |
|
|
148
|
-
| **GPT-5.4** | OpenAI | — | — | ~$2.75 | ✅ | 1M |
|
|
149
|
-
| **Claude Sonnet 4.6** | Anthropic | — | — | ~$3.00 | ✅ | 1M |
|
|
150
|
-
| **Claude Opus 4.8** | Anthropic | — | — | ~$5.00 | ✅ | 1M |
|
|
151
|
-
| **Claude Opus 4.7** | Anthropic | — | — | ~$5.00 | ✅ | 1M |
|
|
152
|
-
| **GPT-5.5** | OpenAI | — | — | ~$5.50 | ✅ | 1M |
|
|
153
|
-
| **Gemini 3.8 Flash Cyber** | Google | — | — | — | — | — |
|
|
154
|
-
|
|
155
|
-
> **Historical footnotes:** The numbered notes below preserve release and pricing context from the prior September 5 snapshot. They are not the current AA ranking source; use the table above and its OpenRouter refresh note for current scores.
|
|
156
|
-
|
|
157
|
-
¹ Gemini 3.1 Pro pricing applies to prompts ≤200K tokens.
|
|
158
|
-
|
|
159
|
-
³ Score is an **estimate** from Artificial Analysis (labelled "independent evaluation forthcoming"). Not a confirmed run of the full evaluation suite. As of August 7, 2026, **GLM 5V Turbo (35.0)** is the only remaining estimate in this table — MiMo V2.5 is now a measured 38.0.
|
|
160
|
-
|
|
161
|
-
⁴ **MiniMax M3 Priority** is the same `MiniMax-M3` weights invoked with `"service_tier": "priority"` in the request body — **not a separate model**. Costs **1.5× Standard** (effective post-50%-off rates: $0.45 / $1.80 / $0.09 per 1M for input/output/cached ≤512K) in exchange for **priority admission** (faster responses, fewer failures during MiniMax peak hours). Capabilities, context window, vision, tools, rate limits, and thinking modes are identical to Standard. See [docs/research/minimax-m3-priority.md](docs/research/minimax-m3-priority.md).
|
|
162
|
-
|
|
163
|
-
⁵ **Claude Sonnet 5** has introductory pricing of **$2.00 / $10.00** per MTok (input/output) through **August 31, 2026**. Standard pricing of $3.00 / $15.00 shown above takes effect September 1, 2026. Released June 30, 2026 — Arena rankings pending. AA Intelligence Index score of **55.3** confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/claude-sonnet-5).
|
|
164
|
-
|
|
165
|
-
⁶ **GPT-5.6** launched July 9, 2026. On July 30, OpenAI reduced Luna pricing by 80% to $0.20/$1.20 and Terra pricing by 20% to $2/$12 per 1M input/output tokens; Sol remains $5/$30. All support image input and 1M context. The Batch API offers an additional 50% discount for asynchronous jobs. OpenRouter is currently running a limited-time 50% promo on Terra and Luna (effective $0.10/$0.60 and $1/$6 per 1M input/output respectively). Benchmark details and cache-write pricing are maintained in [docs/pricing.md](docs/pricing.md).
|
|
166
|
-
|
|
167
|
-
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 **59.7** 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).
|
|
168
|
-
|
|
169
|
-
⁸ **Gemini 3.6 Flash** launched July 21, 2026. **Promotional pricing** of $0.75 / $3.75 per MTok input/output ($0.075 cached) through **December 31, 2026**; $1.50 / $7.50 from January 1, 2027. AA Intelligence Index score of **51.6** confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/gemini-3-6-flash). 1M context with vision.
|
|
170
|
-
|
|
171
|
-
⁹ **Claude Opus 5** launched July 24, 2026. AA Intelligence Index score of **63.1** (#1 overall) confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/claude-opus-5). Priced at $5.00 / $25.00 per MTok input/output (same as Opus 4.8). 1M context, text + image input, adaptive reasoning. Also available in Fast mode ($10/$50 per MTok input/output). Uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). See [Anthropic's Opus 5 announcement](https://www.anthropic.com/news/claude-opus-5).
|
|
172
|
-
|
|
173
|
-
¹⁰ **Qwen 3.8 Max** (`qwen3.8-max`) has a confirmed Artificial Analysis Intelligence Index of **58.1** (Coding **71.8**, Agentic **58.4**) — see [AA](https://artificialanalysis.ai/models/qwen3-8-max) and [OpenRouter](https://openrouter.ai/qwen/qwen3.8-max). Arena snapshots place it **#5 in Text**, **#2 in Vision**, and **#4 in Code/WebDev**; it is not listed in Agent Arena. See [Arena](https://arena.ai/leaderboard).
|
|
174
|
-
|
|
175
|
-
¹¹ **Grok 4.5** (xAI, released July 8, 2026) is now a **GitHub Copilot native** model (GA, Versatile). AA Intelligence Index **55.8** (high), Coding **72.4**, Agentic **48.9**. 500K context, text + image input. Priced at $2.00 / $0.50 / $6.00 per 1M input/cached/output (75% cache discount; Copilot long-context >200K tier $4.00 / $1.00 / $12.00). See the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
176
|
-
|
|
177
|
-
¹² **Qwen 3.7 price cuts** (verified August 7, 2026): `qwen3.7-max` dropped from $2.50 / $7.50 to **$1.475 / $4.425** per 1M input/output (cached $0.295), and `qwen3.7-plus` dropped from $0.40 / $1.60 to **$0.32 / $1.28** per 1M (cached $0.064).
|
|
178
|
-
|
|
179
|
-
¹³ **Gemini 3.7 Flash** (Google, released August 13, 2026) is GitHub Copilot native (GA, Versatile). AA Intelligence Index **56** (high preset) confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/gemini-3-7-flash). **Promotional pricing** of $0.75 / $3.75 per MTok input/output ($0.075 cached) through December 31, 2026, then $1.50 / $7.50. 1M context, text + image + speech + video input, and the fastest model on AA's leaderboard (340 t/s).
|
|
180
|
-
|
|
181
|
-
¹⁴ **Grok 4.6** (xAI, released August 12, 2026) is now a **GitHub Copilot native** model (GA). AA Intelligence Index **61** (high preset, #6/188). 500K context, text + image input. Priced at $2.00 / $0.50 / $6.00 per 1M input/cached/output. See the [AA model page](https://artificialanalysis.ai/models/grok-4-6), the [xAI announcement](https://x.ai/news/grok-4-6), and the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
182
|
-
|
|
183
|
-
¹⁵ **Claude Fable 5** (Anthropic, released June 9, 2026) is now a **GitHub Copilot native** model (GA, Powerful). AA Intelligence Index **62.1** (#3/188) confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/claude-fable-5). **#1 in all four Arena leaderboards** (Text, Agent, Code, Overall). 1M context, text + image input, adaptive reasoning. Priced at $10.00 / $50.00 per MTok input/output ($1.00 cached, 90% cache discount). Uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). **Superseded as Anthropic's flagship by Claude Fable 5.1 (September 2026 — see footnote ²¹)**. See the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
184
|
-
|
|
185
|
-
¹⁶ **Qwen 3.6 Plus** (DashScope, released April 2, 2026) — AA Intelligence Index **40.0**. Priced at $0.50 / $3.00 per MTok input/output ($0.05 cached). 1M context, text + image + video input. **Deprecated** in favor of Qwen 3.7 Plus.
|
|
186
|
-
|
|
187
|
-
¹⁷ **GLM 5.3** (Z.ai, released August 18, 2026) — new flagship. AA Intelligence Index **59.5** (OpenRouter AA-sourced; AA page lists rounded **60**, #8/182). Priced identically to GLM 5.2: $1.40 / $0.26 / $4.40 per 1M input/cached/output. 1M context, text-only, always-thinking with `reasoning_effort` (low/high/max, default max). Released too recently for Arena rankings. See the [AA model page](https://artificialanalysis.ai/models/glm-5-3) and [Z.ai docs](https://docs.z.ai/guides/llm/glm-5.3).
|
|
188
|
-
|
|
189
|
-
¹⁸ **GLM 5.3 Flash** (Z.ai, released August 26, 2026) — Z.ai's first native multimodal GLM-5 model (text + image input, text output). Hybrid sparse + linear attention architecture, 320B total / 18B active parameters (open weights, MIT), 1M context window. AA Intelligence Index **57.5** (Coding **71.5**, Agentic **58.2**). Priced at **$0.15 / $0.03 / $0.50** per 1M input/cached/output (list; 50% off launch promo through September 9, 2026). Always-thinking with `reasoning_effort` (low/high/max, default max). The cheapest cost-per-intelligence row in this table (~$0.0022, ~$0.13/session). See [docs/models/glm.md](docs/models/glm.md) and the [AA model page](https://artificialanalysis.ai/models/glm-5-3-flash).
|
|
190
|
-
|
|
191
|
-
¹⁹ **Gemini 3.8 Flash** (`gemini-3.8-flash`, released September 2, 2026) — Google's most intelligent Flash model, built for long-horizon software engineering and autonomous agents. AA Intelligence Index **58.7** (OpenRouter AA-sourced; AA page rounds to **59**, #17/196), Coding **76.3**, Agentic **50.0**. 1M context, text + image + speech + video + PDF input. Configurable thinking (low/medium/high). **Not yet GitHub Copilot native** — priced from the Gemini API at the same promotional rates as 3.7 Flash ($0.75 / $3.75 per MTok input/output, $0.075 cached) through December 31, 2026, then $1.50 / $7.50. Fast (305.5 t/s) but very verbose — budget more output tokens than the score alone suggests. See the [Google announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/) and the [AA model page](https://artificialanalysis.ai/models/gemini-3-8-flash).
|
|
192
|
-
|
|
193
|
-
²⁰ **Gemini 3.8 Flash Cyber** (released September 2, 2026) — cybersecurity-specialized variant of 3.8 Flash (autonomous vulnerability discovery + automated patching). **Restricted access** via Google's [Fairwind Program](https://deepmind.google/fairwind-program/) — no public API pricing and no AA Intelligence Index run, hence the `—` cells. Vendor-reported CyberGym Pass@1 **86.2** and CWE-Bench Pass@1 **47.2**. See the [Google announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/).
|
|
194
|
-
|
|
195
|
-
²¹ **Claude Fable 5.1** (released September 1, 2026) — Anthropic's new flagship; AA Intelligence Index **65.7** (OpenRouter AA-sourced; AA page rounds to **66**, #1/196), Coding **81.6**, Agentic **61.3** — now the top-scoring model in this table. GitHub Copilot native (GA, Powerful). 1M context, text + image input, adaptive reasoning. Priced at $10.00 / $50.00 per MTok input/output ($0.25 cached — down from Fable 5's $1.00; 97.5% cache discount; $12.50 cache write). Uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). Too new for Arena rankings. See the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) and the [AA model page](https://artificialanalysis.ai/models/claude-fable-5-1).
|
|
196
|
-
|
|
197
|
-
²² **GPT-6 Astra** (released September 3, 2026) — OpenAI's frontier model, available as `gpt-6-astra` through the API and as a GitHub Copilot native model (GA, Powerful). AA Intelligence Index **61.2** (launch comparison, methodology v4.1.1); 1M context, text + image input, text output, and reasoning. Standard pricing is $10.00 / $1.00 / $50.00 per 1M input/cached/output tokens, for a modeled ~$10.00 session and ~$0.163 CPI. It is too new for Arena rankings. See the [GPT-6 Astra announcement](https://openai.com/index/gpt-6-astra/), [OpenAI API pricing](https://developers.openai.com/api/docs/pricing), and [OpenAI model documentation](https://developers.openai.com/api/docs/models/gpt-6-astra).
|
|
198
|
-
|
|
199
|
-
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).
|
|
108
|
+
All prices are **USD per 1M tokens** (non-cached). 1 AI credit = $0.01. To convert to AI credits, multiply by 100 (e.g., $5.00/1M = 500 credits/1M). Session cost assumes ~10K input + ~2K output tokens per turn, 50 turns. DeepSeek V4 estimates use peak rates; official off-peak rates are half. The current AA values come from OpenRouter's September 10 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 (#2)**, while the exact 0902 API metadata row is **40.3**. 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. Current corrections include OpenAI GPT-5.6 Sol at **$4 / $0.40 / $20** and Anthropic Sonnet 5 at **$2 / $0.20 / $10**.
|
|
109
|
+
|
|
110
|
+
| Model | Provider | Cost per intelligence | Intelligence Score | Est. session | Vision | Context window |
|
|
111
|
+
| -------------------------- | --------- | --------------------- | ------------------ | ------------ | ------ | -------------- |
|
|
112
|
+
| **GLM 5.3 Flash** | Z.ai | **~$0.0030** | **41.9** | ~$0.13 | ✅ | 1M |
|
|
113
|
+
| **MiMo V2.5** | Xiaomi | **~$0.0045** | **22.3** | ~$0.10 | ✅ | 1M |
|
|
114
|
+
| **GPT-5.6 Luna** | OpenAI | **~$0.0059** | **37.5** | ~$0.22 | ✅ | 1M |
|
|
115
|
+
| **MiniMax M3** | MiniMax | **~$0.0091** | **29.6** | ~$0.27 | ✅ | 1M |
|
|
116
|
+
| **DeepSeek V4 Flash 0731** | DeepSeek | **~$0.0101** | **34.5** | ~$0.35 | ❌ | 1M |
|
|
117
|
+
| **MiMo V2.5 Pro** | Xiaomi | **~$0.0114** | **26.4** | ~$0.30 | ❌ | 1M |
|
|
118
|
+
| **MiniMax M3 Priority** | MiniMax | **~$0.0139** | **29.6** | ~$0.41 | ✅ | 1M |
|
|
119
|
+
| **Gemini 3.8 Flash** | Google | **~$0.0182** | **41.2** | ~$0.75 | ✅ | 1M |
|
|
120
|
+
| **GLM 5.3** | Z.ai | **~$0.0254** | **44.9** | ~$1.14 | ❌ | 1M |
|
|
121
|
+
| **DeepSeek V4 Pro 0813** | DeepSeek | **~$0.0292** | **36.3** | ~$1.06 | ❌ | 1M |
|
|
122
|
+
| **Qwen 3.8 Max (0803)** | DashScope | **~$0.0300** | **53.4** | ~$1.60 | ✅ | 1M |
|
|
123
|
+
| **Grok 4.6** | xAI | **~$0.0360** | **44.4** | ~$1.60 | ✅ | 500K |
|
|
124
|
+
| **Qwen 3.8 Max (0902)** | DashScope | **~$0.0397** | **40.3** | ~$1.60 | ✅ | 1M |
|
|
125
|
+
| **GPT-5.6 Terra** | OpenAI | **~$0.0520** | **42.3** | ~$2.20 | ✅ | 1M |
|
|
126
|
+
| **Claude Sonnet 5** | Anthropic | **~$0.0521** | **38.4** | ~$2.00 | ✅ | 1M |
|
|
127
|
+
| **Kimi K3** | Moonshot | **~$0.0685** | **43.8** | ~$3.00 | ✅ | 1M |
|
|
128
|
+
| **GPT-5.6 Sol** | OpenAI | **~$0.0849** | **47.1** | ~$4.00 | ✅ | 1M |
|
|
129
|
+
| **Claude Opus 5** | Anthropic | **~$0.0986** | **50.7** | ~$5.00 | ✅ | 1M |
|
|
130
|
+
| **Claude Fable 5.1** | Anthropic | **~$0.1873** | **53.4** | ~$10.00 | ✅ | 1M |
|
|
131
|
+
| **GPT-6 Astra** | OpenAI | **~$0.1894** | **52.8** | ~$10.00 | ✅ | 1M |
|
|
132
|
+
| **Claude Fable 5** | Anthropic | **~$0.2012** | **49.7** | ~$10.00 | ✅ | 1M |
|
|
133
|
+
|
|
134
|
+
> Detailed pricing, benchmark provenance, and model-specific notes live in [docs/pricing.md](docs/pricing.md) and [docs/benchmarks.md](docs/benchmarks.md). This page stays focused on setup and model selection. For a copy-paste config containing **all providers at once**, see [docs/example-config.md](docs/example-config.md).
|
|
200
135
|
|
|
201
136
|
> **👤 Personal picks** —
|
|
202
137
|
>
|
|
203
|
-
>
|
|
204
|
-
>
|
|
205
|
-
> - **GLM 5.3 Flash** — current AA **46.2**, ~$0.13/session, vision-capable, and the best low-cost scored option in this table.
|
|
206
|
-
> - **GPT-5.6 Luna** — current AA **43.4**, ~$0.22/session, vision-capable, and the cheapest current OpenAI scored row.
|
|
207
|
-
> - **DeepSeek V4 Flash 0731** — current AA **40.8**, ~$0.35/session at peak direct rates, text-only and inexpensive for coding-heavy work.
|
|
208
|
-
> - **GPT-6 Astra** — current OpenRouter AA **54.7**, ~$10.00/session, while OpenAI reports leading results on Terminal-Bench 4.0, ARC-AGI-3, FrontierMath Tier 4, and AutomationBench. Use it for the hardest planning and research tasks.
|
|
209
|
-
> - **Qwen 3.8 Max (0902)** — current AA **46.9**, ~$1.60/session, 1M context, and vision-capable; a practical multimodal planning choice.
|
|
210
|
-
>
|
|
211
|
-
> **MiniMax M3** (current AA **35.7**, ~$0.27/session) remains the simplest direct custom-endpoint option: no proxy or extension, vision, 1M context, and tool calling.
|
|
138
|
+
> - **GLM 5.3 Flash** — the lowest cost per intelligence in this table (~$0.0030), at about $0.13/session with vision.
|
|
139
|
+
> - **GPT-5.6 Luna** — the lowest cost per intelligence among Copilot-native scored models (~$0.0059), at about $0.22/session with vision.
|
|
212
140
|
|
|
213
141
|
## Companion tools
|
|
214
142
|
|
|
@@ -223,7 +151,7 @@ VS Code's built-in `view_image` tool only accepts **static images** (PNG, JPG, G
|
|
|
223
151
|
**Video Context MCP** is a small MCP server that bridges that gap. It works with **GitHub Copilot, Cursor, and Claude Code** out of the box, and:
|
|
224
152
|
|
|
225
153
|
- **Extracts frames** from local files or remote URLs (no `ffmpeg` gymnastics required).
|
|
226
|
-
- **Routes them through a multi-provider fallback chain** — `Gemini → GLM 4.6V Flash → Qwen3.
|
|
154
|
+
- **Routes them through a multi-provider fallback chain** — `Gemini → GLM 4.6V Flash → Qwen3.8-max → MiMo-V2.5`.
|
|
227
155
|
- **Answers natural-language questions** about the video grounded in actual frames: "what does the speaker click in the last 30 seconds?", "summarize the demo", "find the frame where the error appears".
|
|
228
156
|
- **Extras:** timestamp search, audio transcription with speaker diarization, and video metadata (resolution, duration, codec).
|
|
229
157
|
|
package/docs/example-config.md
CHANGED
|
@@ -6,7 +6,7 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
|
|
|
6
6
|
>
|
|
7
7
|
> **Note:** The `apiKey` fields are left as empty strings — set them via the **Chat: Manage Language Models** UI (Command Palette → right-click provider group → **Update API Key**). After you set a key via the UI, VS Code replaces the empty string with a `${input:chat.lm.secret.<id>}` secret reference.
|
|
8
8
|
>
|
|
9
|
-
> The live config points Qwen at the local proxy (`:3458`) and MiMo at the local proxy (`:3459`). When using a proxy, align the model `requestBody` overrides with the proxy's behavior: Qwen sends no `requestBody` (the proxy manages `enable_thinking` dynamically); MiMo sends only `temperature` and `top_p` (the proxy injects `thinking: {"type": "disabled"}` on tool turns and leaves it absent on plain chat).
|
|
9
|
+
> The live config points Qwen at the local proxy (`:3458`) and MiMo at the local proxy (`:3459`). When using a proxy, align the model `requestBody` overrides with the proxy's behavior: Qwen sends no `requestBody` (the proxy manages `enable_thinking` dynamically); MiMo sends only `temperature` and `top_p` (the proxy injects `thinking: {"type": "disabled"}` on tool turns and leaves it absent on plain chat). OpenRouter lists Qwen 3.8 Max as two frozen snapshots (`qwen/qwen3.8-max` = 0803 and `qwen/qwen3.8-max-0902` = 0902); this DashScope config intentionally uses the provider ID `qwen3.8-max` and does not pretend the OpenRouter slugs are interchangeable.
|
|
10
10
|
|
|
11
11
|
```json
|
|
12
12
|
[
|
|
@@ -25,26 +25,6 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
|
|
|
25
25
|
"streaming": true,
|
|
26
26
|
"maxInputTokens": 991000,
|
|
27
27
|
"maxOutputTokens": 131072
|
|
28
|
-
},
|
|
29
|
-
{
|
|
30
|
-
"id": "qwen3.7-max",
|
|
31
|
-
"name": "Qwen 3.7 Max (text)",
|
|
32
|
-
"url": "http://127.0.0.1:3458/v1/chat/completions",
|
|
33
|
-
"toolCalling": true,
|
|
34
|
-
"vision": false,
|
|
35
|
-
"streaming": true,
|
|
36
|
-
"maxInputTokens": 1048576,
|
|
37
|
-
"maxOutputTokens": 131072
|
|
38
|
-
},
|
|
39
|
-
{
|
|
40
|
-
"id": "qwen3.7-plus",
|
|
41
|
-
"name": "Qwen 3.7 Plus (vision)",
|
|
42
|
-
"url": "http://127.0.0.1:3458/v1/chat/completions",
|
|
43
|
-
"toolCalling": true,
|
|
44
|
-
"vision": true,
|
|
45
|
-
"streaming": true,
|
|
46
|
-
"maxInputTokens": 1048576,
|
|
47
|
-
"maxOutputTokens": 131072
|
|
48
28
|
}
|
|
49
29
|
]
|
|
50
30
|
},
|
|
@@ -67,33 +47,6 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
|
|
|
67
47
|
"streaming": true,
|
|
68
48
|
"maxInputTokens": 1000000,
|
|
69
49
|
"maxOutputTokens": 131072
|
|
70
|
-
},
|
|
71
|
-
{
|
|
72
|
-
"id": "kimi-k2.6",
|
|
73
|
-
"name": "Kimi K2.6 (vision)",
|
|
74
|
-
"url": "http://127.0.0.1:3457/v1/chat/completions",
|
|
75
|
-
"requestBody": {
|
|
76
|
-
"temperature": 1
|
|
77
|
-
},
|
|
78
|
-
"toolCalling": true,
|
|
79
|
-
"vision": true,
|
|
80
|
-
"streaming": true,
|
|
81
|
-
"maxInputTokens": 262144,
|
|
82
|
-
"maxOutputTokens": 32768
|
|
83
|
-
},
|
|
84
|
-
{
|
|
85
|
-
"id": "kimi-k2.7-code",
|
|
86
|
-
"name": "Kimi K2.7 Code (vision)",
|
|
87
|
-
"url": "http://127.0.0.1:3457/v1/chat/completions",
|
|
88
|
-
"requestBody": {
|
|
89
|
-
"temperature": 1,
|
|
90
|
-
"max_tokens": 4096
|
|
91
|
-
},
|
|
92
|
-
"toolCalling": true,
|
|
93
|
-
"vision": true,
|
|
94
|
-
"streaming": true,
|
|
95
|
-
"maxInputTokens": 262144,
|
|
96
|
-
"maxOutputTokens": 4096
|
|
97
50
|
}
|
|
98
51
|
]
|
|
99
52
|
},
|
|
@@ -194,51 +147,6 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
|
|
|
194
147
|
"temperature": 1.0,
|
|
195
148
|
"top_p": 0.95
|
|
196
149
|
}
|
|
197
|
-
},
|
|
198
|
-
{
|
|
199
|
-
"id": "glm-5.2",
|
|
200
|
-
"name": "GLM 5.2 (text)",
|
|
201
|
-
"url": "https://api.z.ai/api/paas/v4/chat/completions",
|
|
202
|
-
"toolCalling": true,
|
|
203
|
-
"vision": false,
|
|
204
|
-
"streaming": true,
|
|
205
|
-
"maxInputTokens": 1048576,
|
|
206
|
-
"maxOutputTokens": 131072,
|
|
207
|
-
"requestBody": {
|
|
208
|
-
"thinking": { "type": "enabled" },
|
|
209
|
-
"temperature": 1.0,
|
|
210
|
-
"top_p": 0.95
|
|
211
|
-
}
|
|
212
|
-
},
|
|
213
|
-
{
|
|
214
|
-
"id": "glm-5.1",
|
|
215
|
-
"name": "GLM 5.1 (text)",
|
|
216
|
-
"url": "https://api.z.ai/api/paas/v4/chat/completions",
|
|
217
|
-
"toolCalling": true,
|
|
218
|
-
"vision": false,
|
|
219
|
-
"streaming": true,
|
|
220
|
-
"maxInputTokens": 204800,
|
|
221
|
-
"maxOutputTokens": 131072,
|
|
222
|
-
"requestBody": {
|
|
223
|
-
"thinking": { "type": "enabled" },
|
|
224
|
-
"temperature": 1,
|
|
225
|
-
"top_p": 0.95
|
|
226
|
-
}
|
|
227
|
-
},
|
|
228
|
-
{
|
|
229
|
-
"id": "glm-5v-turbo",
|
|
230
|
-
"name": "GLM 5V Turbo (vision)",
|
|
231
|
-
"url": "https://api.z.ai/api/paas/v4/chat/completions",
|
|
232
|
-
"toolCalling": true,
|
|
233
|
-
"vision": true,
|
|
234
|
-
"streaming": true,
|
|
235
|
-
"maxInputTokens": 204800,
|
|
236
|
-
"maxOutputTokens": 131072,
|
|
237
|
-
"requestBody": {
|
|
238
|
-
"thinking": { "type": "enabled" },
|
|
239
|
-
"temperature": 1,
|
|
240
|
-
"top_p": 0.95
|
|
241
|
-
}
|
|
242
150
|
}
|
|
243
151
|
]
|
|
244
152
|
}
|
|
@@ -250,9 +158,9 @@ Here's a complete, real-world `chatLanguageModels.json` that combines **the `cus
|
|
|
250
158
|
If you only need one provider, jump straight to its setup guide:
|
|
251
159
|
|
|
252
160
|
- [Kimi K3 / K2.6 / K2.7 Code](kimi.md)
|
|
253
|
-
- [Qwen 3.
|
|
161
|
+
- [Qwen 3.8 Max](qwen.md)
|
|
254
162
|
- [Xiaomi MiMo (V2.5 / V2.5 Pro)](mimo.md)
|
|
255
163
|
- [MiniMax M3 (Standard)](minimax.md) and [MiniMax M3 (Priority)](minimax.md#5-m3-priority-tier-optional)
|
|
256
|
-
- [GLM (5.3 Flash / 5.3 / 5.2 / 5.1
|
|
164
|
+
- [GLM (5.3 Flash / 5.3 / 5.2 / 5.1)](glm.md)
|
|
257
165
|
|
|
258
166
|
> **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.
|
package/docs/models/glm.md
CHANGED
|
@@ -18,8 +18,8 @@
|
|
|
18
18
|
| ---------------------- | ------------------------------------------------------------------------------------------------------------------ |
|
|
19
19
|
| Mode | **Direct** (no proxy) |
|
|
20
20
|
| Billing | **Pay-as-You-Go** (PaaS API) — Coding Plan subscription exists but is **not usable** from VS Code custom endpoints |
|
|
21
|
-
| Vision | ✅ Yes (`glm-5.3-flash`
|
|
22
|
-
| Tool calling | ✅ Yes (native multimodal tool use on `glm-5.3-flash`
|
|
21
|
+
| Vision | ✅ Yes (`glm-5.3-flash`) |
|
|
22
|
+
| Tool calling | ✅ Yes (native multimodal tool use on `glm-5.3-flash`) |
|
|
23
23
|
| Context | 1M (`glm-5.3-flash` / `glm-5.3` / `glm-5.2` Solid Lossless Context) |
|
|
24
24
|
| Max output | 131072 |
|
|
25
25
|
| Required `requestBody` | `thinking: { type: "enabled" }` + `reasoning_effort: "max"` (recommended) |
|
|
@@ -35,7 +35,6 @@
|
|
|
35
35
|
| `glm-5.3` | ❌ | 1M | 131072 | **New flagship** — ~50% coding gain over 5.2 on Z.ai Code Bench; always-thinking, `reasoning_effort` `low`/`high`/`max` (default `max`) ($1.40 / $4.40 per 1M) |
|
|
36
36
|
| `glm-5.2` | ❌ | 1M | 131072 | Previous flagship — "Opus-level" long-context engineering & agentic coding; thinking `enabled` ($1.40 / $4.40 per 1M) |
|
|
37
37
|
| `glm-5.1` | ❌ | 200K | 131072 | Previous flagship — long-horizon / 8h autonomous work; thinking `enabled` ($1.40 / $4.40 per 1M) |
|
|
38
|
-
| `glm-5v-turbo` | ✅ | 200K | 131072 | Multimodal **coding** model — vision-based agentic coding; thinking `enabled` ($1.20 / $4.00 per 1M) |
|
|
39
38
|
|
|
40
39
|
> Other GLM models (`glm-5`, `glm-5-turbo`, `glm-4.5-air`, etc.) are callable on the same endpoint but are intentionally **not** added to the default `chatLanguageModels.json` block below. Add them in the same shape if you need them. Note: `glm-4.6v-flashx` was previously in the default block but has been **removed** because live testing showed it is not reliable for tool calling.
|
|
41
40
|
|
|
@@ -96,51 +95,6 @@
|
|
|
96
95
|
"temperature": 1.0,
|
|
97
96
|
"top_p": 0.95
|
|
98
97
|
}
|
|
99
|
-
},
|
|
100
|
-
{
|
|
101
|
-
"id": "glm-5.2",
|
|
102
|
-
"name": "GLM 5.2 (text)",
|
|
103
|
-
"url": "https://api.z.ai/api/paas/v4/chat/completions",
|
|
104
|
-
"toolCalling": true,
|
|
105
|
-
"vision": false,
|
|
106
|
-
"streaming": true,
|
|
107
|
-
"maxInputTokens": 1048576,
|
|
108
|
-
"maxOutputTokens": 131072,
|
|
109
|
-
"requestBody": {
|
|
110
|
-
"thinking": { "type": "enabled" },
|
|
111
|
-
"temperature": 1.0,
|
|
112
|
-
"top_p": 0.95
|
|
113
|
-
}
|
|
114
|
-
},
|
|
115
|
-
{
|
|
116
|
-
"id": "glm-5.1",
|
|
117
|
-
"name": "GLM 5.1 (text)",
|
|
118
|
-
"url": "https://api.z.ai/api/paas/v4/chat/completions",
|
|
119
|
-
"toolCalling": true,
|
|
120
|
-
"vision": false,
|
|
121
|
-
"streaming": true,
|
|
122
|
-
"maxInputTokens": 204800,
|
|
123
|
-
"maxOutputTokens": 131072,
|
|
124
|
-
"requestBody": {
|
|
125
|
-
"thinking": { "type": "enabled" },
|
|
126
|
-
"temperature": 1,
|
|
127
|
-
"top_p": 0.95
|
|
128
|
-
}
|
|
129
|
-
},
|
|
130
|
-
{
|
|
131
|
-
"id": "glm-5v-turbo",
|
|
132
|
-
"name": "GLM 5V Turbo (vision)",
|
|
133
|
-
"url": "https://api.z.ai/api/paas/v4/chat/completions",
|
|
134
|
-
"toolCalling": true,
|
|
135
|
-
"vision": true,
|
|
136
|
-
"streaming": true,
|
|
137
|
-
"maxInputTokens": 204800,
|
|
138
|
-
"maxOutputTokens": 131072,
|
|
139
|
-
"requestBody": {
|
|
140
|
-
"thinking": { "type": "enabled" },
|
|
141
|
-
"temperature": 1,
|
|
142
|
-
"top_p": 0.95
|
|
143
|
-
}
|
|
144
98
|
}
|
|
145
99
|
]
|
|
146
100
|
}
|
|
@@ -169,7 +123,7 @@
|
|
|
169
123
|
- **`tool_choice` only supports `auto`** — don't override it (VS Code's default is `auto`).
|
|
170
124
|
- **`clear_thinking` defaults to `true` server-side**, which strips historical `reasoning_content` from prior turns before sending to the model. This is what makes multi-turn tool loops stable. **Do not** set `clear_thinking: false` from `requestBody` — VS Code can't forward `reasoning_content` and the request will fail. (Note: Z.ai's GLM-5.3-Flash docs recommend `thinking.clear_thinking: false` for clients that **do** preserve `reasoning_content` — e.g. Claude Code / agent SDKs — not for the VS Code custom-endpoint path.)
|
|
171
125
|
- **GLM 5.3 and GLM 5.3 Flash are always-thinking** — `thinking.type` only supports `enabled`; disabling reasoning is **not supported**. `reasoning_effort` accepts `low`, `high`, or `max` (default `max`; `max` recommended for coding). If you previously sent `thinking.type: "disabled"` for other GLM models, switch it to `enabled` with `reasoning_effort: "low"` when moving to `glm-5.3` / `glm-5.3-flash`.
|
|
172
|
-
- **Vision** is supported on `glm-5.3-flash` (native multimodal — the first in the GLM-5 series; text + image input, text output)
|
|
126
|
+
- **Vision** is supported on `glm-5.3-flash` (native multimodal — the first in the GLM-5 series; text + image input, text output). `glm-5.3`, `glm-5.2` and `glm-5.1` are text-only. For turnkey VS Code video understanding, see [**Video Context MCP**](https://www.videocontextmcp.com/). **Validated 2026-08-28 on `glm-5.3-flash`:** both vision surfaces work through the custom endpoint — chat image attachments (base64 `image_url`) and the `view_image` tool on workspace files. Fine-grained text, layout, and image content were read correctly in both tests.
|
|
173
127
|
- **Context caching** is automatic — `usage.prompt_tokens_details.cached_tokens` reports cache hits; cache writes are currently free. Cache reads for `glm-5.3-flash` are $0.03/M (list; $0.015/M during the launch promo).
|
|
174
128
|
|
|
175
129
|
## Troubleshooting
|
package/docs/models/kimi.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# Kimi (Moonshot) — VS Code Custom Endpoint Setup Guide
|
|
2
2
|
|
|
3
|
-
> **TL;DR:** Kimi
|
|
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.
|
|
4
4
|
>
|
|
5
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.
|
|
6
6
|
>
|
|
@@ -22,13 +22,11 @@
|
|
|
22
22
|
|
|
23
23
|
### Models
|
|
24
24
|
|
|
25
|
-
| Model
|
|
26
|
-
|
|
|
27
|
-
| `kimi-k3`
|
|
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). |
|
|
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**. |
|
|
30
28
|
|
|
31
|
-
> Deprecated K2 ids (`kimi-k2-0905-preview`, `kimi-k2-turbo-preview`, `kimi-k2-thinking`, etc.) were discontinued May 25, 2026
|
|
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.
|
|
32
30
|
|
|
33
31
|
## Quick Start
|
|
34
32
|
|
|
@@ -36,7 +34,7 @@
|
|
|
36
34
|
2. **Edit `chatLanguageModels.json`** — add the Kimi block below.
|
|
37
35
|
3. **Set your Moonshot API key** via Command Palette → **Chat: Manage Language Models**.
|
|
38
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 V2.5). [Why?](../../README.md#4-configure-the-utility-small-model)
|
|
39
|
-
5. **Restart VS Code** and pick "Kimi K3"
|
|
37
|
+
5. **Restart VS Code** and pick "Kimi K3".
|
|
40
38
|
|
|
41
39
|
## Setup
|
|
42
40
|
|
|
@@ -65,36 +63,12 @@
|
|
|
65
63
|
"streaming": true,
|
|
66
64
|
"maxInputTokens": 1000000,
|
|
67
65
|
"maxOutputTokens": 131072
|
|
68
|
-
},
|
|
69
|
-
{
|
|
70
|
-
"id": "kimi-k2.6",
|
|
71
|
-
"name": "Kimi K2.6 (vision)",
|
|
72
|
-
"url": "http://127.0.0.1:3457/v1/chat/completions",
|
|
73
|
-
"requestBody": { "temperature": 1 },
|
|
74
|
-
"toolCalling": true,
|
|
75
|
-
"vision": true,
|
|
76
|
-
"streaming": true,
|
|
77
|
-
"maxInputTokens": 262144,
|
|
78
|
-
"maxOutputTokens": 32768
|
|
79
|
-
},
|
|
80
|
-
{
|
|
81
|
-
"id": "kimi-k2.7-code",
|
|
82
|
-
"name": "Kimi K2.7 Code (vision)",
|
|
83
|
-
"url": "http://127.0.0.1:3457/v1/chat/completions",
|
|
84
|
-
"requestBody": { "temperature": 1, "max_tokens": 4096 },
|
|
85
|
-
"toolCalling": true,
|
|
86
|
-
"vision": true,
|
|
87
|
-
"streaming": true,
|
|
88
|
-
"maxInputTokens": 262144,
|
|
89
|
-
"maxOutputTokens": 4096
|
|
90
66
|
}
|
|
91
67
|
]
|
|
92
68
|
}
|
|
93
69
|
```
|
|
94
70
|
|
|
95
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.
|
|
96
|
-
>
|
|
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.
|
|
98
72
|
|
|
99
73
|
### 2. API key
|
|
100
74
|
|
|
@@ -138,7 +112,6 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
|
|
|
138
112
|
- Forwards your `Authorization` header upstream.
|
|
139
113
|
- Rewrites plain-chat requests to `temperature: 1`, `top_p: 0.95`.
|
|
140
114
|
- For **K2.5 / K2.6**: rewrites tool-enabled requests to `thinking: { type: "disabled" }`, `temperature: 0.6`, `top_p: 0.95`.
|
|
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
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.
|
|
143
116
|
- Preserves streaming responses (SSE).
|
|
144
117
|
- Writes redacted request summaries to `debug_log/kimi-proxy.ndjson`.
|
|
@@ -146,10 +119,9 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
|
|
|
146
119
|
## Notes
|
|
147
120
|
|
|
148
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).
|
|
149
|
-
- **
|
|
150
|
-
-
|
|
151
|
-
-
|
|
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.
|
|
122
|
+
- **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.
|
|
153
125
|
|
|
154
126
|
## Troubleshooting
|
|
155
127
|
|
|
@@ -167,10 +139,8 @@ Set in `.env` at the repo root (the proxy `import 'dotenv/config'` automatically
|
|
|
167
139
|
|
|
168
140
|
For the cross-provider comparison, see [docs/pricing.md](../pricing.md). Moonshot direct platform rates:
|
|
169
141
|
|
|
170
|
-
| Model
|
|
171
|
-
|
|
|
172
|
-
| `kimi-k3`
|
|
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 |
|
|
142
|
+
| Model | Input (cache miss) | Cached input | Output |
|
|
143
|
+
| --------- | ------------------ | ------------ | ----------- |
|
|
144
|
+
| `kimi-k3` | $3.00 / 1M | $0.30 / 1M | $15.00 / 1M |
|
|
175
145
|
|
|
176
|
-
> K3 pricing is flat (no tiering by context length).
|
|
146
|
+
> K3 pricing is flat (no tiering by context length).
|
package/docs/models/qwen.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# Qwen (DashScope) — VS Code Custom Endpoint Setup Guide
|
|
2
2
|
|
|
3
|
-
> **TL;DR:** `qwen3.8-max` (vision)
|
|
3
|
+
> **TL;DR:** `qwen3.8-max` (vision) works both direct and via the local proxy. OpenRouter's September 10 rankings card shows unversioned Qwen3.8 Max at **#2 with 53.4**, while the exact `qwen/qwen3.8-max-0902` API metadata row is **40.3**; the public API no longer exposes the unversioned ID. The proxy forwards the upstream `model` value unchanged, so use the exact model ID supported by the endpoint you choose. The proxy gives you dynamic thinking suppression: reasoning stays ON in plain chat but turns OFF automatically when tools are invoked. The direct path is simpler if you don't need reasoning in chat.
|
|
4
4
|
|
|
5
5
|
## At a Glance
|
|
6
6
|
|
|
@@ -8,10 +8,10 @@
|
|
|
8
8
|
| ---------------------- | -------------------------------------------------------------------------------- |
|
|
9
9
|
| Mode | **Proxy** (local on `:3458`) **or** **Direct** (static `enable_thinking: false`) |
|
|
10
10
|
| Billing | **Pay-as-You-Go only** — 1M-token free quota for new users |
|
|
11
|
-
| Vision | ✅ Yes (`qwen3.8-max
|
|
11
|
+
| Vision | ✅ Yes (`qwen3.8-max`) |
|
|
12
12
|
| Tool calling | ✅ Yes |
|
|
13
13
|
| Context | 1M |
|
|
14
|
-
| Max output | 131K
|
|
14
|
+
| Max output | 131K |
|
|
15
15
|
| Endpoint | `https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions` |
|
|
16
16
|
| Proxy endpoint | `http://127.0.0.1:3458/v1/chat/completions` |
|
|
17
17
|
| Auth | `Authorization: Bearer $DASHSCOPE_API_KEY` |
|
|
@@ -20,14 +20,74 @@
|
|
|
20
20
|
|
|
21
21
|
### Models
|
|
22
22
|
|
|
23
|
-
| Model
|
|
24
|
-
|
|
|
25
|
-
| `qwen3.8-max`
|
|
26
|
-
| `qwen3.7-plus` | ✅ Yes | 1M | _(unspecified)_ | Primary model with image understanding |
|
|
27
|
-
| `qwen3.7-max` | ❌ No | 1M | _(unspecified)_ | Larger text-only model |
|
|
23
|
+
| Model | Vision | Context | Max output | Notes |
|
|
24
|
+
| ------------- | ------ | ------- | ---------- | ------------------------------------------------------- |
|
|
25
|
+
| `qwen3.8-max` | ✅ Yes | 1M | 131K | Frontier multimodal model; reasoning enabled by default |
|
|
28
26
|
|
|
29
27
|
> The live `chatLanguageModels.json` points Qwen at the local proxy by default; the direct DashScope URL is shown below for users who prefer a static `enable_thinking: false` setup.
|
|
30
28
|
|
|
29
|
+
### OpenRouter snapshot records
|
|
30
|
+
|
|
31
|
+
These are separate OpenRouter catalog entries, not two picker labels for one model:
|
|
32
|
+
|
|
33
|
+
| OpenRouter slug | Snapshot | AA Intelligence Index | Notes |
|
|
34
|
+
| ----------------------- | -------- | --------------------- | ---------------------------------------------------------------------------------- |
|
|
35
|
+
| `qwen/qwen3.8-max` | 0803 | **53.4 (#2)** | September 10 ranking-card result; exact unversioned API entry is no longer exposed |
|
|
36
|
+
| `qwen/qwen3.8-max-0902` | 0902 | **40.3** | September snapshot; current exact-ID API metadata |
|
|
37
|
+
|
|
38
|
+
The DashScope custom-endpoint snippets below use the provider model ID `qwen3.8-max`. Do not replace it with the OpenRouter `qwen/qwen3.8-max-0902` slug unless you also change the upstream URL to OpenRouter and have verified that path independently.
|
|
39
|
+
|
|
40
|
+
### Optional OpenRouter snapshot proxy
|
|
41
|
+
|
|
42
|
+
The proxy forwards the `model` field without mapping it, so it can also be pointed at OpenRouter for the two catalog snapshots. This is an **optional, unvalidated OpenRouter path**; it is separate from the validated DashScope setup above and requires an OpenRouter API key.
|
|
43
|
+
|
|
44
|
+
1. Add this to the repo-root `.env` file:
|
|
45
|
+
|
|
46
|
+
```dotenv
|
|
47
|
+
QWEN_UPSTREAM_URL=https://openrouter.ai/api/v1/chat/completions
|
|
48
|
+
```
|
|
49
|
+
|
|
50
|
+
2. Start the existing Qwen proxy:
|
|
51
|
+
|
|
52
|
+
```bash
|
|
53
|
+
npm run proxy:qwen
|
|
54
|
+
```
|
|
55
|
+
|
|
56
|
+
3. Add a separate provider group using the OpenRouter key and these model entries:
|
|
57
|
+
|
|
58
|
+
```json
|
|
59
|
+
{
|
|
60
|
+
"name": "Qwen (OpenRouter snapshots)",
|
|
61
|
+
"vendor": "customendpoint",
|
|
62
|
+
"apiKey": "",
|
|
63
|
+
"apiType": "chat-completions",
|
|
64
|
+
"models": [
|
|
65
|
+
{
|
|
66
|
+
"id": "qwen/qwen3.8-max",
|
|
67
|
+
"name": "Qwen 3.8 Max (0803)",
|
|
68
|
+
"url": "http://127.0.0.1:3458/v1/chat/completions",
|
|
69
|
+
"toolCalling": true,
|
|
70
|
+
"vision": true,
|
|
71
|
+
"streaming": true,
|
|
72
|
+
"maxInputTokens": 991000,
|
|
73
|
+
"maxOutputTokens": 131072
|
|
74
|
+
},
|
|
75
|
+
{
|
|
76
|
+
"id": "qwen/qwen3.8-max-0902",
|
|
77
|
+
"name": "Qwen 3.8 Max (0902)",
|
|
78
|
+
"url": "http://127.0.0.1:3458/v1/chat/completions",
|
|
79
|
+
"toolCalling": true,
|
|
80
|
+
"vision": true,
|
|
81
|
+
"streaming": true,
|
|
82
|
+
"maxInputTokens": 991000,
|
|
83
|
+
"maxOutputTokens": 131072
|
|
84
|
+
}
|
|
85
|
+
]
|
|
86
|
+
}
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
Set the API key through **Chat: Manage Language Models**, then verify the proxy with `curl http://127.0.0.1:3458/healthz`. Do not run this route at the same time as the default DashScope route unless each process uses a different `QWEN_PROXY_PORT` and log path.
|
|
90
|
+
|
|
31
91
|
## Quick Start
|
|
32
92
|
|
|
33
93
|
### With Proxy (recommended)
|
|
@@ -43,7 +103,7 @@
|
|
|
43
103
|
1. **Use the direct-path JSON snippet** below.
|
|
44
104
|
2. **Set your `DASHSCOPE_API_KEY`** via Command Palette → **Chat: Manage Language Models**.
|
|
45
105
|
3. **Configure the Utility Small Model** — Open Settings → search **"Chat: Utility Small Model"** → pick your fastest model (e.g., DeepSeek V4 Flash or MiMo V2.5). [Why?](../../README.md#4-configure-the-utility-small-model)
|
|
46
|
-
4. **Restart VS Code** and pick "Qwen 3.8 Max"
|
|
106
|
+
4. **Restart VS Code** and pick "Qwen 3.8 Max".
|
|
47
107
|
|
|
48
108
|
## Setup
|
|
49
109
|
|
|
@@ -84,24 +144,6 @@ DashScope is region-specific — your API key only works on the endpoint it was
|
|
|
84
144
|
"maxInputTokens": 991000,
|
|
85
145
|
"maxOutputTokens": 131072,
|
|
86
146
|
"requestBody": { "enable_thinking": false }
|
|
87
|
-
},
|
|
88
|
-
{
|
|
89
|
-
"id": "qwen3.7-max",
|
|
90
|
-
"name": "Qwen 3.7 Max (text)",
|
|
91
|
-
"url": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions",
|
|
92
|
-
"toolCalling": true,
|
|
93
|
-
"vision": false,
|
|
94
|
-
"streaming": true,
|
|
95
|
-
"requestBody": { "enable_thinking": false }
|
|
96
|
-
},
|
|
97
|
-
{
|
|
98
|
-
"id": "qwen3.7-plus",
|
|
99
|
-
"name": "Qwen 3.7 Plus (vision)",
|
|
100
|
-
"url": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions",
|
|
101
|
-
"toolCalling": true,
|
|
102
|
-
"vision": true,
|
|
103
|
-
"streaming": true,
|
|
104
|
-
"requestBody": { "enable_thinking": false }
|
|
105
147
|
}
|
|
106
148
|
]
|
|
107
149
|
}
|
|
@@ -141,22 +183,6 @@ DashScope is region-specific — your API key only works on the endpoint it was
|
|
|
141
183
|
"streaming": true,
|
|
142
184
|
"maxInputTokens": 991000,
|
|
143
185
|
"maxOutputTokens": 131072
|
|
144
|
-
},
|
|
145
|
-
{
|
|
146
|
-
"id": "qwen3.7-max",
|
|
147
|
-
"name": "Qwen 3.7 Max (text)",
|
|
148
|
-
"url": "http://127.0.0.1:3458/v1/chat/completions",
|
|
149
|
-
"toolCalling": true,
|
|
150
|
-
"vision": false,
|
|
151
|
-
"streaming": true
|
|
152
|
-
},
|
|
153
|
-
{
|
|
154
|
-
"id": "qwen3.7-plus",
|
|
155
|
-
"name": "Qwen 3.7 Plus (vision)",
|
|
156
|
-
"url": "http://127.0.0.1:3458/v1/chat/completions",
|
|
157
|
-
"toolCalling": true,
|
|
158
|
-
"vision": true,
|
|
159
|
-
"streaming": true
|
|
160
186
|
}
|
|
161
187
|
]
|
|
162
188
|
}
|
|
@@ -217,8 +243,8 @@ The proxy detects active tool use by examining the conversation state (a `"tool"
|
|
|
217
243
|
|
|
218
244
|
## Notes
|
|
219
245
|
|
|
220
|
-
- **Vision (`qwen3.8-max
|
|
221
|
-
- **Qwen 3.8 reasoning:** `qwen3.8-max` enables reasoning by default and supports `reasoning_effort` values `low`, `medium`, and `xhigh` (default). The direct snippet disables reasoning for stable VS Code tool loops; the proxy leaves it on for plain chat and suppresses it when tool activity is detected.
|
|
246
|
+
- **Vision (`qwen3.8-max`)** uses OpenAI-compatible `content` array format. Base64 data URIs work reliably; external image URLs may fail if DashScope can't reach them. If a drag-and-drop image fails to load, providing the absolute file path (e.g. `c:\path\to\image.png`) in the prompt is a reliable workaround.
|
|
247
|
+
- **Qwen 3.8 reasoning:** `qwen3.8-max` enables reasoning by default and supports `reasoning_effort` values `low`, `medium`, and `xhigh` (default). The direct snippet disables reasoning for stable VS Code tool loops; the proxy leaves it on for plain chat and suppresses it when tool activity is detected. The OpenRouter ranking card and 0902 API metadata are separate source records; the benchmark difference does not mean the DashScope config should silently switch IDs.
|
|
222
248
|
- **Thinking trade-off:** Direct = thinking always off (loops stable, no reasoning visible). Proxy = thinking on in plain chat, off in tool turns.
|
|
223
249
|
- **`tool_choice` only supports `auto`** — don't override it (VS Code's default is `auto`).
|
|
224
250
|
|
|
@@ -237,12 +263,10 @@ The proxy detects active tool use by examining the conversation state (a `"tool"
|
|
|
237
263
|
|
|
238
264
|
For the cross-provider comparison, see [docs/pricing.md](../pricing.md). DashScope (international) rates for **non-thinking** mode:
|
|
239
265
|
|
|
240
|
-
| Model
|
|
241
|
-
|
|
|
242
|
-
| `qwen3.8-max`
|
|
243
|
-
| `qwen3.7-plus` | $0.32 / 1M | $0.40 / 1M | $1.28 / 1M | $1.60 / 1M |
|
|
244
|
-
| `qwen3.7-max` | $1.475 / 1M (≤ 1M) | — | $4.425 / 1M (≤ 1M) | — |
|
|
266
|
+
| Model | Input (≤ 256K tokens) | Input (> 256K tokens) | Output (≤ 256K tokens) | Output (> 256K tokens) |
|
|
267
|
+
| ------------- | --------------------- | --------------------- | ---------------------- | ---------------------- |
|
|
268
|
+
| `qwen3.8-max` | $2.00 | — | $6.00 | — |
|
|
245
269
|
|
|
246
|
-
> **Qwen
|
|
270
|
+
> **Billing options:** Qwen Cloud now offers a separate Token Plan subscription with access to Qwen3.8 Max. The validated VS Code snippets in this document use the PAYG DashScope-compatible endpoint and API key; do not assume the Token Plan base URL or credentials are interchangeable without following Qwen Cloud's current integration instructions.
|
|
247
271
|
|
|
248
272
|
> Free quota: 1M input + 1M output tokens per model, valid 90 days after activating Model Studio.
|
package/docs/pricing.md
CHANGED
|
@@ -1,16 +1,14 @@
|
|
|
1
1
|
# Pricing
|
|
2
2
|
|
|
3
|
-
>
|
|
3
|
+
> Release history is kept in the reference notes below. Use the current table and September 10 refresh note for present scores and estimates. The September 11 roster cleanup removes two retired models without changing the remaining pricing data.
|
|
4
4
|
|
|
5
|
-
> **
|
|
6
|
-
|
|
7
|
-
> **Current refresh (September 6, 2026):** AA Intelligence Index values below use the current OpenRouter model pages. **Every pricing cell is checked against first-party provider pricing or GitHub Copilot's official billing table; OpenRouter is never used as a pricing source.** Qwen 3.8 Max is now represented by the current **0902** snapshot. Several models no longer expose a current OpenRouter AA composite and are shown as `—` rather than retaining stale scores. Current corrections include OpenAI GPT-5.6 Sol at **$4 / $0.40 / $20**, Anthropic Sonnet 5 at **$2 / $0.20 / $10**, xAI Grok 4.5 cache reads at **$0.30**, and Alibaba's Qwen 3.7 list rates at **$2.50 / $0.25 / $7.50** (Max) and **$0.40 / $0.04 / $1.60** (Plus).
|
|
5
|
+
> **Latest refresh (September 10, 2026):** OpenRouter's live benchmark card puts the unversioned Qwen3.8 Max entry at **#2 with 53.4**. The public model API no longer exposes that unversioned ID and canonicalizes it to 0902; the exact `qwen/qwen3.8-max-0902` metadata row currently reports **40.3**. The table preserves both values as separate records rather than silently merging them. Qwen Cloud's current PAYG API rates remain **$2 / $0.25 implicit cache / $6** for Qwen3.8 Max; Qwen Cloud also now offers a separate Token Plan subscription. GLM 5.3 Flash's 50% launch promotion ended September 9, so list rates are now the applicable baseline.
|
|
8
6
|
|
|
9
7
|
> **⏰ June 1, 2026 — GitHub Copilot switched to usage-based billing (AI Credits) today.**
|
|
10
8
|
>
|
|
11
|
-
> Before this change, Copilot used **premium request-based billing** — each model had its own multiplier
|
|
9
|
+
> 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.
|
|
12
10
|
>
|
|
13
|
-
> **The practical workaround:** use cheaper alternative models (DeepSeek V4 Flash,
|
|
11
|
+
> **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.
|
|
14
12
|
>
|
|
15
13
|
> 1 AI credit = $0.01 USD. All paid plans include a monthly credit allowance:
|
|
16
14
|
>
|
|
@@ -26,56 +24,31 @@ All prices below are in **USD per 1M tokens** (non-cached). To convert to AI cre
|
|
|
26
24
|
|
|
27
25
|
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.
|
|
28
26
|
|
|
29
|
-
| Model
|
|
30
|
-
|
|
|
31
|
-
| **GLM 5.3 Flash**
|
|
32
|
-
| **
|
|
33
|
-
| **
|
|
34
|
-
| **
|
|
35
|
-
| **
|
|
36
|
-
| **
|
|
37
|
-
| **
|
|
38
|
-
| **Gemini 3.
|
|
39
|
-
| **
|
|
40
|
-
| **
|
|
41
|
-
| **
|
|
42
|
-
| **Grok 4.6**
|
|
43
|
-
| **Qwen 3.8 Max (0902)
|
|
44
|
-
| **
|
|
45
|
-
| **Claude Sonnet 5**
|
|
46
|
-
| **
|
|
47
|
-
| **
|
|
48
|
-
| **
|
|
49
|
-
| **
|
|
50
|
-
| **
|
|
51
|
-
| **Claude Fable 5
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
| **MiMo V2.5** | Xiaomi | — | — | $0.14 | $0.0028 | $0.28 | ~$0.10 | 1M |
|
|
55
|
-
| **Qwen 3.7 Plus** | DashScope | — | — | $0.40 (≤256K) / $1.20 (>256K) | $0.04 | $1.60 (≤256K) / $4.80 (>256K) | ~$0.36 | 1M |
|
|
56
|
-
| **Qwen 3.6 Plus** | DashScope | — | — | $0.50 | $0.05 | $3.00 | ~$0.55 | 1M |
|
|
57
|
-
| **GPT-5.4 mini** | OpenAI | — | — | $0.75 | $0.075 | $4.50 | ~$0.83 | 400K |
|
|
58
|
-
| **Kimi K2.6** | Moonshot | — | — | $0.95 | $0.16 | $4.00 | ~$0.88 | 256K |
|
|
59
|
-
| **Kimi K2.7 Code** | Moonshot | — | — | $0.95 | $0.19 | $4.00 | ~$0.88 | 262K |
|
|
60
|
-
| **GLM 5V Turbo** | Z.ai | — | — | $1.20 | $0.24 | $4.00 | ~$1.00 | 200K |
|
|
61
|
-
| **GLM 5.1** | Z.ai | — | — | $1.40 | $0.26 | $4.40 | ~$1.14 | 200K |
|
|
62
|
-
| **GLM 5.2** | Z.ai | — | — | $1.40 | $0.26 | $4.40 | ~$1.14 | 1M |
|
|
63
|
-
| **Gemini 3.5 Flash** | Google | — | — | $1.50 | $0.15 | $9.00 | ~$1.65 | 1M |
|
|
64
|
-
| **Qwen 3.7 Max** | DashScope | — | — | $2.50 (≤1M) | $0.25 | $7.50 (≤1M) | ~$2.00 | 1M |
|
|
65
|
-
| **GPT-5.4** | OpenAI | — | — | $2.50 | $0.25 | $15.00 | ~$2.75 | 1M |
|
|
66
|
-
| **Claude Sonnet 4.6** | Anthropic | — | — | $3.00 | $0.30 | $15.00 | ~$3.00 | 1M |
|
|
67
|
-
| **Claude Opus 4.8** | Anthropic | — | — | $5.00 | $0.50 | $25.00 | ~$5.00 | 1M |
|
|
68
|
-
| **Claude Opus 4.7** | Anthropic | — | — | $5.00 | $0.50 | $25.00 | ~$5.00 | 1M |
|
|
69
|
-
| **GPT-5.5** | OpenAI | — | — | $5.00 | $0.50 | $30.00 | ~$5.50 | 1M |
|
|
70
|
-
| **Gemini 3.8 Flash Cyber** | Google | — | — | — | — | — | — | — |
|
|
71
|
-
|
|
72
|
-
> **Historical footnotes:** The numbered notes below preserve release and pricing context from the prior September 5 snapshot. They are not the current AA ranking source; use the table above and the September 6 OpenRouter refresh note for current scores.
|
|
73
|
-
|
|
74
|
-
¹ Gemini 3.1 Pro pricing applies to prompts ≤200K tokens.
|
|
75
|
-
|
|
76
|
-
³ Score is an **estimate** from Artificial Analysis (labelled "independent evaluation forthcoming"). Not a confirmed run of the full evaluation suite. As of August 7, 2026, **GLM 5V Turbo (35.0)** is the only remaining estimate in this table — MiMo V2.5 is now a measured 38.0.
|
|
77
|
-
|
|
78
|
-
⁴ GLM 5.2 was Z.ai's previous flagship (superseded by **GLM 5.3** on August 18, 2026 — see footnote ¹⁸). AA Intelligence Index score (**52.6**) and pricing of **$1.40** (input), **$0.26** (cached), and **$4.40** (output) per 1M tokens confirmed on its [model page](https://artificialanalysis.ai/models/glm-5-2). Session cost is identical to GLM 5.1 (~$1.14). As of August 17, 2026, OpenRouter's cheapest provider is **Open Sail Research** at **$0.50 / $3.15** per 1M (listed price, no promo); NovitaAI's temporary promo has been reduced from 52% to **48% off** (effective **$0.7238 / $2.275** per 1M, cache reads $0.1344).
|
|
27
|
+
| Model | Provider | Cost per intelligence | Intelligence Score | Input (per 1M) | Cached input | Output (per 1M) | Est. session | Context window |
|
|
28
|
+
| -------------------------- | --------- | --------------------- | ------------------ | ----------------------------- | ----------------------------- | ----------------------------- | ------------ | -------------- |
|
|
29
|
+
| **GLM 5.3 Flash** | Z.ai | **~$0.0030** | **41.9** | $0.15 | $0.03 | $0.50 | ~$0.13 | 1M |
|
|
30
|
+
| **MiMo V2.5** | Xiaomi | **~$0.0045** | **22.3** | $0.14 | $0.0028 | $0.28 | ~$0.10 | 1M |
|
|
31
|
+
| **GPT-5.6 Luna** | OpenAI | **~$0.0059** | **37.5** | $0.20 | $0.02 | $1.20 | ~$0.22 | 1M |
|
|
32
|
+
| **MiniMax M3** | MiniMax | **~$0.0091** | **29.6** | $0.60 (≤512K) / $1.20 (>512K) | $0.12 (≤512K) / $0.24 (>512K) | $2.40 (≤512K) / $4.80 (>512K) | ~$0.27 | 1M |
|
|
33
|
+
| **DeepSeek V4 Flash 0731** | DeepSeek | **~$0.0101** | **34.5** | $0.44 | $0.014 | $1.32 | ~$0.35 | 1M |
|
|
34
|
+
| **MiMo V2.5 Pro** | Xiaomi | **~$0.0114** | **26.4** | $0.435 | $0.0036 | $0.87 | ~$0.30 | 1M |
|
|
35
|
+
| **MiniMax M3 Priority** | MiniMax | **~$0.0139** | **29.6** | $0.90 (≤512K) / $1.80 (>512K) | $0.18 (≤512K) / $0.36 (>512K) | $3.60 (≤512K) / $7.20 (>512K) | ~$0.41 | 1M |
|
|
36
|
+
| **Gemini 3.8 Flash** | Google | **~$0.0182** | **41.2** | $0.75 | $0.075 | $3.75 | ~$0.75 | 1M |
|
|
37
|
+
| **GLM 5.3** | Z.ai | **~$0.0254** | **44.9** | $1.40 | $0.26 | $4.40 | ~$1.14 | 1M |
|
|
38
|
+
| **DeepSeek V4 Pro 0813** | DeepSeek | **~$0.0292** | **36.3** | $1.32 | $0.044 | $3.96 | ~$1.06 | 1M |
|
|
39
|
+
| **Qwen 3.8 Max (0803)²⁵** | DashScope | **~$0.0300** | **53.4** | $2.00 | $0.25 | $6.00 | ~$1.60 | 1M |
|
|
40
|
+
| **Grok 4.6** | xAI | **~$0.0360** | **44.4** | $2.00 | $0.50 | $6.00 | ~$1.60 | 500K |
|
|
41
|
+
| **Qwen 3.8 Max (0902)²⁶** | DashScope | **~$0.0397** | **40.3** | $2.00 | $0.25 | $6.00 | ~$1.60 | 1M |
|
|
42
|
+
| **GPT-5.6 Terra** | OpenAI | **~$0.0520** | **42.3** | $2.00 | $0.20 | $12.00 | ~$2.20 | 1M |
|
|
43
|
+
| **Claude Sonnet 5** | Anthropic | **~$0.0521** | **38.4** | $2.00 | $0.20 | $10.00 | ~$2.00 | 1M |
|
|
44
|
+
| **Kimi K3** | Moonshot | **~$0.0685** | **43.8** | $3.00 | $0.30 | $15.00 | ~$3.00 | 1M |
|
|
45
|
+
| **GPT-5.6 Sol** | OpenAI | **~$0.0849** | **47.1** | $4.00 | $0.40 | $20.00 | ~$4.00 | 1M |
|
|
46
|
+
| **Claude Opus 5** | Anthropic | **~$0.0986** | **50.7** | $5.00 | $0.50 | $25.00 | ~$5.00 | 1M |
|
|
47
|
+
| **Claude Fable 5.1** | Anthropic | **~$0.1873** | **53.4** | $10.00 | $0.25 | $50.00 | ~$10.00 | 1M |
|
|
48
|
+
| **GPT-6 Astra** | OpenAI | **~$0.1894** | **52.8** | $10.00 | $1.00 | $50.00 | ~$10.00 | 1M |
|
|
49
|
+
| **Claude Fable 5** | Anthropic | **~$0.2012** | **49.7** | $10.00 | $1.00 | $50.00 | ~$10.00 | 1M |
|
|
50
|
+
|
|
51
|
+
> **Reference notes:** The details below preserve release and pricing context. Use the current table and September 10 refresh note for present values.
|
|
79
52
|
|
|
80
53
|
⁵ **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 (list prices shown above; effective rates after the standing 50% off are $0.45/$1.80/$0.09 ≤512K and $0.90/$3.60/$0.18 >512K), 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. To enable it on the custom-endpoint entry, add `"service_tier": "priority"` to the `requestBody` of the single `MiniMax-M3` block (and remove it to go back to Standard). See [docs/models/minimax.md](models/minimax.md#4-m3-priority-tier-optional) and [docs/research/minimax-m3-priority.md](research/minimax-m3-priority.md) for the full breakdown.
|
|
81
54
|
|
|
@@ -85,25 +58,15 @@ All models are listed together below, sorted by Cost per intelligence ascending
|
|
|
85
58
|
|
|
86
59
|
⁸ **GPT-5.6** pricing is **$0.20 / $0.02 / $1.20** for Luna, **$2.00 / $0.20 / $12.00** for Terra, and **$5.00 / $0.50 / $30.00** for Sol (input / cached input / output per 1M tokens). OpenAI reduced Luna by 80% and Terra by 20% on July 30, 2026; Sol is unchanged. OpenAI bills cache writes at 1.25x the uncached input rate for GPT-5.6 and later. The Batch API provides an additional 50% discount for asynchronous jobs. All three tiers have a 1M-token context window and support text + image input. See [OpenAI's GPT-5.6 price update](https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/) and [GPT-5.6 announcement](https://openai.com/index/gpt-5-6/). **As of August 7, 2026, OpenRouter is running a limited-time 50% promotional discount on Terra and Luna** (effective **$0.10 / $0.60** per 1M for Luna and **$1.00 / $6.00** per 1M for Terra, input/output) on top of the list prices above — this is a temporary third-party promo, not an OpenAI list-price change.
|
|
87
60
|
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
¹⁰ **Claude Opus 5** was released on **July 24, 2026**. AA Intelligence Index score (**63.1**) confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/claude-opus-5). Pricing is $5.00 / $0.50 / $25.00 per MTok input/cached/output (same as Opus 4.8). 1M context window, text + image input, adaptive reasoning. Also available in Fast mode ($10/$50 per MTok input/output) and Batch ($2.50/$12.50 per MTok input/output). Uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). See [Anthropic's Opus 5 announcement](https://www.anthropic.com/news/claude-opus-5).
|
|
91
|
-
|
|
92
|
-
¹¹ **Qwen 3.8 Max** (`qwen3.8-max`) launched on **August 3, 2026**. Qwen Cloud lists $2.00 input / $6.00 output per 1M tokens, $0.25 implicit-cache input, $2.50 explicit-cache creation, and $0.17 explicit-cache reads. The modeled session cost is ~$1.60 using 10K input + 2K output tokens per turn across 50 turns. It has a 1M context window, 131K maximum output, and text/image/video input. Artificial Analysis has published an Intelligence Index of **58.1** for this model (**#9/185**; listed as a rounded **58** on its [model page](https://artificialanalysis.ai/models/qwen3-8-max)); OpenRouter metadata also reports Coding **71.8** and Agentic **58.4**. Pricing is unchanged as of August 7, 2026. See the [Qwen Cloud model page](https://www.qwencloud.com/models/qwen3.8-max), [AA model page](https://artificialanalysis.ai/models/qwen3-8-max), and [OpenRouter model page](https://openrouter.ai/qwen/qwen3.8-max).
|
|
61
|
+
¹⁰ **Claude Opus 5** was released on **July 24, 2026**. AA Intelligence Index score (**63.1**) confirmed by [Artificial Analysis](https://artificialanalysis.ai/models/claude-opus-5). Pricing is $5.00 / $0.50 / $25.00 per MTok input/cached/output. 1M context window, text + image input, adaptive reasoning. Also available in Fast mode ($10/$50 per MTok input/output) and Batch ($2.50/$12.50 per MTok input/output). Uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). See [Anthropic's Opus 5 announcement](https://www.anthropic.com/news/claude-opus-5).
|
|
93
62
|
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
¹³ **Qwen 3.7 price cuts** (verified August 7, 2026): `qwen3.7-max` dropped from $2.50 / $7.50 to **$1.475 / $4.425** per 1M input/output (cached $0.295; >256K input tier $1.844), and `qwen3.7-plus` dropped from $0.40 / $1.60 to **$0.32 / $1.28** per 1M (cached $0.064; >256K tier $0.40 / $1.60).
|
|
63
|
+
¹¹ **Historical Qwen 3.8 Max launch record:** the August 3, 2026 launch package used the unversioned `qwen3.8-max` name. An earlier Artificial Analysis snapshot reported **58.1**; the current OpenRouter ranking now distinguishes the unversioned **0803** row (**53.4**) from the separate **0902** row (**46.9**). Use the current rows and footnotes ²⁵–²⁶ for comparisons rather than carrying the earlier score forward. See the [Qwen Cloud model page](https://www.qwencloud.com/models/qwen3.8-max), [AA model page](https://artificialanalysis.ai/models/qwen3-8-max), and [OpenRouter model page](https://openrouter.ai/qwen/qwen3.8-max).
|
|
97
64
|
|
|
98
65
|
¹⁴ **Grok 4.6** (xAI/SpaceXAI, released **August 12, 2026**) is now a **GitHub Copilot native** model (GA). AA Intelligence Index **61** (high reasoning preset, #6/188). 500K context window, text + image input, text output. Priced at $2.00 / $0.50 / $6.00 per 1M input/cached/output (75% cache discount). See the [AA model page](https://artificialanalysis.ai/models/grok-4-6), the [xAI announcement](https://x.ai/news/grok-4-6), and the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
99
66
|
|
|
100
|
-
¹⁵ **Gemini 3.7 Flash** (Google, released **August 13, 2026**) — GitHub Copilot native (GA, Versatile) at the same promotional rates as 3.6 Flash. AA Intelligence Index **56** (high reasoning preset, #17/188). **Promotional pricing** of $0.75 / $0.075 / $3.75 per 1M input/cached/output through **December 31, 2026**, then $1.50 / $0.15 / $7.50 from January 1, 2027. 1M context window, text + image + speech + video input, and the fastest model on AA's leaderboard (340 t/s). See the [AA model page](https://artificialanalysis.ai/models/gemini-3-7-flash), [Google's Gemini page](https://deepmind.google/models/gemini/), and [GitHub Copilot pricing](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
101
|
-
|
|
102
67
|
¹⁶ **Claude Fable 5** (Anthropic, released **June 9, 2026**) is now a **GitHub Copilot native** model (GA, Powerful; Anthropic's new flagship tier, above Opus). AA Intelligence Index **62.1** (**#3/188**; AA page lists a rounded **62** on its [model page](https://artificialanalysis.ai/models/claude-fable-5)). Currently **#1 in Text, Agent, Code, and Overall Arena**. 1M context window, text + image input, adaptive reasoning. Priced at **$10.00 / $1.00 / $50.00** per 1M input/cached/output (90% cache discount; cache write $12.50). Uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). **Superseded as Anthropic's flagship by Claude Fable 5.1 (September 1, 2026 — see footnote ²³)**. See the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
103
68
|
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
¹⁸ **GLM 5.3** is Z.ai's newest flagship (released **August 18, 2026**) — same base model as GLM 5.2 with post-training improvements (~50% coding gain over 5.2 on Z.ai Code Bench). AA Intelligence Index **59.5** (OpenRouter's AA-sourced benchmark; AA's [model page](https://artificialanalysis.ai/models/glm-5-3) lists a rounded **60**, **#8/182**), Coding **74.8**, Agentic **59.1**. Pricing is identical to GLM 5.2: **$1.40** input, **$0.26** cached, **$4.40** output per 1M tokens (so the modeled session cost is the same, ~$1.14). 1M context window, 753B params, text-only. Always-thinking: `thinking.type` only supports `enabled` and `reasoning_effort` accepts `low` / `high` / `max` (default `max`). See the [Z.ai model page](https://docs.z.ai/guides/llm/glm-5.3).
|
|
69
|
+
¹⁸ **GLM 5.3** is Z.ai's newest flagship (released **August 18, 2026**) with post-training improvements over the previous GLM flagship (~50% coding gain on Z.ai Code Bench). AA Intelligence Index **59.5** (OpenRouter's AA-sourced benchmark; AA's [model page](https://artificialanalysis.ai/models/glm-5-3) lists a rounded **60**, **#8/182**), Coding **74.8**, Agentic **59.1**. Pricing is **$1.40** input, **$0.26** cached, and **$4.40** output per 1M tokens (modeled session cost ~$1.14). 1M context window, 753B params, text-only. Always-thinking: `thinking.type` only supports `enabled` and `reasoning_effort` accepts `low` / `high` / `max` (default `max`). See the [Z.ai model page](https://docs.z.ai/guides/llm/glm-5.3).
|
|
107
70
|
|
|
108
71
|
¹⁹ **DeepSeek V4 Flash 0731** and **DeepSeek V4 Pro 0813** are the current API builds. Flash's AA Intelligence Index is **51.8** and Pro's is **53.2** (current OpenRouter/Artificial Analysis snapshots). Both are text-only with 1M context; Flash has 284B total / 13B active parameters and Pro has 1.6T total / 49B active. DeepSeek's official V4 pricing changed on August 16, 2026 to peak/off-peak rates: Flash is **$0.44 / $0.014 / $1.32** and Pro is **$1.32 / $0.044 / $3.96** per 1M input/cache-hit/output at peak, with off-peak rates at half those values. See the [DeepSeek pricing page](https://api-docs.deepseek.com/quick_start/pricing), [update log](https://api-docs.deepseek.com/updates/), and [canonical model record](models/deepseek.md).
|
|
109
72
|
|
|
@@ -111,33 +74,34 @@ All models are listed together below, sorted by Cost per intelligence ascending
|
|
|
111
74
|
|
|
112
75
|
²¹ **Gemini 3.8 Flash** (`gemini-3.8-flash`) was released on **September 2, 2026** — Google's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and enterprise workflows. AA Intelligence Index **58.7** (OpenRouter's AA-sourced benchmark; AA's [model page](https://artificialanalysis.ai/models/gemini-3-8-flash) rounds to **59**, **#17/196**), Coding **76.3**, Agentic **50.0**. 1M context window (1,048,576 input / 65,536 output), text + image + speech + video + PDF input, text output; configurable thinking (`low`/`medium`/`high` — `minimal` is not supported). Fast at **305.5 t/s** per AA but **very verbose** (120M output tokens on the Intelligence Index vs a 71M median) with a high TTFT (13.39 s), so expect output-token costs above what the score alone suggests. **Not yet GitHub Copilot native** (GitHub's Google lineup still ends at 3.7 Flash) — priced from the Gemini API at the **same promotional rates as 3.7 Flash**: $0.75 / $0.075 / $3.75 per 1M input/cached/output through **December 31, 2026**, then $1.50 / $0.15 / $7.50 from January 1, 2027 (90% cache discount; Batch API 50% off; cache storage $0.50/1M per hour during the promo). Vendor-reported: HLE-Verified **54.9**. See the [Google announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/), the [Gemini API model page](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash), and the [AA model page](https://artificialanalysis.ai/models/gemini-3-8-flash).
|
|
113
76
|
|
|
114
|
-
²² **Gemini 3.8 Flash Cyber** was released on **September 2, 2026** as a cybersecurity-specialized variant sharing 3.8 Flash's foundational intelligence, focused on autonomous vulnerability discovery and automated patching (defensive capabilities prioritized over offensive ones). It is **not generally available**: access is limited to trusted defenders — government authorities, critical-infrastructure operators, and software maintainers — via Google's [Fairwind Program](https://deepmind.google/fairwind-program/). It has **no public API pricing** and **no Artificial Analysis Intelligence Index run**, so its cells show `—` (listed at the bottom per the table's sort rules) and it is not usable from VS Code custom endpoints. Vendor-reported results: CyberGym Pass@1 **86.2** (ahead of GPT-5.5-Cyber 85.6, Mythos 5 83.8, GPT-5.6 Sol 83.6, and Gemini 3.5 Flash Cyber 77.5), >70% success on Google's internal 20-language vulnerability-discovery benchmark (71.0 vs Gemini 3.7 Flash 58.9 and 3.5 Flash Cyber 46.6), CWE-Bench Pass@1 **47.2** (on the Pareto frontier vs Fable 5's 47.8 at significantly lower cost per rollout), and a **6.0** attack-success rate on the Gray Swan IPI prompt-injection benchmark (lower is better). Real-world reports: 2.6× more correct Chrome patches than larger commercial models; +7.5–9.7% pentest recall at 2.3–5.2× lower cost per Wiz. These are vendor-reported numbers, not independent replications. See the [Google announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/) and the [model page](https://deepmind.google/models/gemini/cyber/).
|
|
115
|
-
|
|
116
77
|
²³ **Claude Fable 5.1** was released on **September 1, 2026** — Anthropic's new flagship, improving on Fable 5 across the board with the biggest gains in agentic coding, long-running agentic workflows, and knowledge work (long code refactors, front-end/visual code generation, finance and analysis), and tending to be more concise in plans and summaries. AA Intelligence Index **65.7** (OpenRouter's AA-sourced benchmark; AA's [model page](https://artificialanalysis.ai/models/claude-fable-5-1) rounds to **66** and ranks it **#1/196**) — the **highest score in this table** — with Coding **81.6** and Agentic **61.3**. GitHub Copilot native (GA, Powerful). 1M context window, text + image input, text output, adaptive reasoning; uses the newer Claude tokenizer (~30% more tokens than pre-4.7 models). Priced at **$10.00 / $0.25 / $50.00** per 1M input/cached/output — the same rates as Fable 5 except **cached input drops from $1.00 to $0.25** (97.5% cache discount); cache write $12.50. Modeled session cost stays ~$10.00, and CPI improves only to ~$0.15 (vs Fable 5's ~$0.161) — still the second-most expensive scored row. At max effort it is slow (66.2 t/s per AA) and very verbose (140M output tokens on the Intelligence Index vs a 71M median; TTFT 283.83 s). Like Fable 5, Anthropic retains request data by default, with zero-data-retention available through the end of 2026 under a time-bound exemption. See the [GitHub Copilot models & pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) and the [AA model page](https://artificialanalysis.ai/models/claude-fable-5-1).
|
|
117
78
|
|
|
118
79
|
²⁴ **GPT-6 Astra** was released on **September 3, 2026** as `gpt-6-astra`, an OpenAI API model and **GitHub Copilot native** model (GA, Powerful). It supports reasoning, text + image input, text output, and a 1M-token context window. Standard pricing is **$10.00 / $1.00 / $50.00** per 1M input/cached/output tokens; cache writes are $12.50/M, and prompts above 272K input tokens use long-context rates. The modeled session cost is ~$10.00 and CPI is ~$0.163. See the [GPT-6 Astra announcement](https://openai.com/index/gpt-6-astra/), [OpenAI API pricing](https://developers.openai.com/api/docs/pricing), and [GitHub Copilot models & pricing](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing).
|
|
119
80
|
|
|
81
|
+
²⁵ **Qwen 3.8 Max (0803)** is the August 3, 2026 launch checkpoint represented by OpenRouter's unversioned ranking card [`qwen/qwen3.8-max`](https://openrouter.ai/qwen/qwen3.8-max). The September 10 ranking card places it at **#2 with 53.4**. It shares Qwen Cloud's **$2.00 / $0.25 implicit cache / $6.00** PAYG rates, 1M context, 131K maximum output, and text/image/video input with the 0902 snapshot. The validated DashScope setup continues to use provider ID `qwen3.8-max` and the proxy forwards that ID unchanged.
|
|
82
|
+
|
|
83
|
+
²⁶ **Qwen 3.8 Max (0902)** is a separate OpenRouter snapshot, [`qwen/qwen3.8-max-0902`](https://openrouter.ai/qwen/qwen3.8-max-0902), released September 4, 2026. The current public model API reports an AA Intelligence Index of **40.3** for this exact ID. It has the same published **$2.00 / $0.25 implicit cache / $6.00** rates, 1M context, 131K maximum output, and text/image/video input. The OpenRouter slug is not automatically a valid DashScope model ID; see [the Qwen setup guide](models/qwen.md) before changing the upstream URL.
|
|
84
|
+
|
|
120
85
|
Cost per intelligence = estimated session cost ÷ Intelligence Index score. Session cost assumes ~10K input + ~2K output tokens per turn, 50 turns.
|
|
121
86
|
|
|
122
87
|
> **Notes:**
|
|
123
88
|
>
|
|
124
89
|
> - **DeepSeek V4** input pricing shown is the **peak cache miss** price. Peak cache hits are $0.014/M for Flash and $0.044/M for Pro; off-peak rates are half. See footnote ¹⁹ for the full schedule.
|
|
125
90
|
> - **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.
|
|
126
|
-
> - **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
|
|
91
|
+
> - **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 5, Sonnet 5, and Fable 5 / 5.1 use a new tokenizer that produces approximately 30% more tokens for the same text.
|
|
127
92
|
> - **OpenAI** models support cached input at 0.1× base input rate.
|
|
128
93
|
> - **Qwen** models use **tiered pricing** — determined by total input tokens per request. Prices above are for non-thinking mode.
|
|
129
94
|
> - **Kimi** official tables list **Cache Hit before Cache Miss** (opposite order to our table). The rows below transpose them so "Input" = cache miss and "Cached input" = cache hit.
|
|
130
|
-
> - **Kimi K2.6** pricing is from the **Moonshot platform** (direct). Via DashScope: $0.89 input / $3.71 output.
|
|
131
95
|
> - **DashScope** offers a **free quota** of 1M input + 1M output tokens per model, valid for 90 days.
|
|
132
96
|
> - **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.
|
|
133
97
|
> - **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.
|
|
134
|
-
> - **GLM** models support prompt caching — cache hits are priced at $0.26/M for 5.3
|
|
98
|
+
> - **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).
|
|
135
99
|
> - **MiMo** offers a **Token Plan** subscription model with discounted rates and a free cache-writing promotion.
|
|
136
100
|
> - For typical Copilot chat usage (short-to-medium prompts), you'll almost always fall in the lowest pricing tier.
|
|
137
101
|
|
|
138
|
-
> **How long does 7,000 credits last?** A Pro+ subscriber running 50-turn sessions
|
|
102
|
+
> **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.
|
|
139
103
|
|
|
140
|
-
> Prices last verified: September
|
|
104
|
+
> Prices last verified: September 10, 2026. Always check the official pages for the latest rates:
|
|
141
105
|
>
|
|
142
106
|
> - [GitHub Copilot models & pricing](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing)
|
|
143
107
|
> - [OpenAI pricing](https://openai.com/api/pricing/)
|
package/package.json
CHANGED
package/proxy/qwen-proxy.mjs
CHANGED
|
@@ -5,7 +5,10 @@ import { createProxy } from '../lib/create-proxy.mjs'
|
|
|
5
5
|
|
|
6
6
|
/**
|
|
7
7
|
* Supported model scope for this proxy:
|
|
8
|
-
* - Validated with `qwen3.8-max
|
|
8
|
+
* - Validated with `qwen3.8-max`.
|
|
9
|
+
* - The rewrite can also be used with OpenRouter model IDs such as
|
|
10
|
+
* `qwen/qwen3.8-max-0902`; the proxy forwards model IDs unchanged. That route
|
|
11
|
+
* is not part of the validated DashScope setup.
|
|
9
12
|
* - Expected to work for any Qwen3 hybrid-thinking model (qwen3-* series)
|
|
10
13
|
* that supports the `enable_thinking` top-level field on DashScope's
|
|
11
14
|
* OpenAI-compatible surface.
|
|
@@ -33,9 +36,13 @@ Starts a local HTTP proxy that conditionally injects enable_thinking: false
|
|
|
33
36
|
when the request includes a tools array, letting Qwen hybrid-thinking models
|
|
34
37
|
show reasoning in plain chat while keeping tool loops stable.
|
|
35
38
|
|
|
39
|
+
The proxy forwards the requested model ID unchanged, including the DashScope
|
|
40
|
+
ID qwen3.8-max and OpenRouter ID qwen/qwen3.8-max-0902.
|
|
41
|
+
|
|
36
42
|
Environment variables:
|
|
37
43
|
QWEN_PROXY_PORT Local listen port. Default: 3458 (falls back to PORT)
|
|
38
|
-
QWEN_UPSTREAM_URL Upstream
|
|
44
|
+
QWEN_UPSTREAM_URL Upstream chat-completions URL (DashScope by default;
|
|
45
|
+
OpenRouter can be selected for snapshot IDs).
|
|
39
46
|
Default: https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions
|
|
40
47
|
QWEN_PROXY_DISABLE_THINKING_WITH_TOOLS
|
|
41
48
|
Inject enable_thinking: false when tools are present.
|