Kimi K3.1 Pro: API pricing and availability
Kimi K3.1 Pro: API pricing and availability
Kimi K3.1 Pro is available through Kimi and its API. Full weights are announced for release by November 4, 2026. I checked the public pricing on October 23: the standard model costs $2 per million input tokens and $10 per million output tokens. Highspeed costs $4 and $20 respectively. The claimed 2-3x speed increase comes from Moonshot; I have not measured it.
| Model | Input / 1M tokens | Output / 1M tokens | Context |
|---|---|---|---|
| K3.1 Pro | 2 USD | 10 USD | 1 048 576 |
| Highspeed | 4 USD | 20 USD | 1 048 576 |
Official sources
- Moonshot: Kimi K3.1 Pro - Read: 2026-10-23
- Kimi API - Read: 2026-10-23
- API: tokens - Read: 2026-10-23
Benchmarks reported by Moonshot
According to Moonshot, 2026-10-23. These are vendor results, not my measurements. Different tests are not interchangeable.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 66.4 | 58.2 |
| SWE Marathon v1.1 | 51.7 | 46.3 | 42.5 |
| HLE-Full | 51.1 | 59.8 | 53.6 |
| HLE-Full w/ tools | 70.0 | 67.7 | 57.2 |
| OSWorld V2.1 | 79.6 | 81.8 | 75.8 |
| GDPval-AA v2.1 (Elo-score) | 1846 | 1866 | 1573 |
| OfficeQA Pro v2 | 64.4 | 62.2 | 54.4 |
| Frontier Finance | 58.6 | 56.7 | 45.0 |
Independent Artificial Analysis Intelligence Index: 55.1 (2026-10-24).
Kimi K3 from China's Moonshot AI debuted on July 16, 2026, and a week and a half later the company released the full weights - and it was precisely this move that triggered sanctions discussions in Washington. I'm gathering in one place the pricing, benchmarks, and how to use the model. Only verified facts, updated in real time.
as of: 11 November 2026
Key facts
- Premiera 16 July 2026, open weights on Hugging Face from 27 July (1,56 TB, MXFP4)
- 2.8 trillion parameters (104B active per token), context 1M tokens, image support - the world's largest open model
- API: 3 / 15 USD per million tokens (input / output), input cache $0.30; subscriptions from $19/month
- 1st place in Frontend Code Arena (1679 pts, ahead of Fable 5 and GPT-5.6 Sol); Artificial Analysis: 57.1 pts, 4th place out of 189 models
- Weak point: hallucinations 50.9% in AA-Omniscience (predecessor K2.6: 39.3%)
- Moonshot AI valued at $31.5 billion (up from $20 billion in May 2026)
Articles about Kimi K3
Frequently Asked Questions
What is Kimi K3?
Kimi K3 is a large language model from China's Moonshot AI, released July 16, 2026. Mixture-of-Experts architecture: 2.8 trillion parameters total, 104 billion active per token, context window of 1,048,576 tokens, supports text and images. It's the world's largest open-weight model.
How much does Kimi K3 cost?
API: $3 per million input tokens, $15 per million output, and cache hits only $0.30. The model always "thinks" and reasoning tokens are billed as output. Kimi app subscriptions: $19, $39, $99, and $199 monthly.
Is Kimi K3 free?
Weights are open: since July 27, 2026 they're on Hugging Face (moonshotai/Kimi-K3, 1.56 TB in MXFP4 format). Running it yourself isn't a laptop project though - Moonshot recommends a cluster with at least 64 accelerators and about 1.4 TB total GPU memory. For most people, the realistic path is API or the app.
Is Kimi K3 better than Claude and GPT?
Depends on the task. In the Frontend Code Arena ranking, K3 leads (1679 pts) ahead of Claude Fable 5 (1631) and GPT-5.6 Sol (1618). In Artificial Analysis's overall index it scores 57.1 points and fourth place among 189 models. Moonshot itself admits that overall K3 still trails the strongest closed models, and AA tests showed a higher hallucination rate than its predecessor (50.9 versus 39.3 percent in AA-Omniscience).
Who is Moonshot AI?
This is a Chinese AI startup, creator of the Kimi model family (K2, K2.5, K2.6, K2.7-Code) and the Kimi.com app. According to the Financial Times, the company is raising funding at a $31.5 billion valuation (up from $20 billion in May 2026).
// follow
New tests and price changes, straight to your feed reader
Free RSS, no account and no email needed.
Follow Quantslant →Sources: TechCrunch (via Financial Times) · Arena - leaderboard · Hugging Face: moonshotai · Moonshot AI
British English (UK) test: 70.5/72 against Opus 5.5 and GPT-6 Astra
In my 12-task British English API test, Kimi K3.1 Pro scored 70.5/72 (4th of 7), effectively tying Claude Opus 5.5 (71.5) and GPT-6 Astra (71.0) while dropping just 1.5 points on a forced pun and brief code comments. These results reflect a single OpenRouter trial on the moonshotai/mxfp4 endpoint and cannot be generalised to the Kimi web app or other languages.
| Model | Score | Tasks |
|---|---|---|
| Claude Opus 5.5 | 71.5 / 72 | 12 |
| DeepSeek V4.1 Flash 1017 | 71 / 72 | 12 |
| GPT-6 Astra | 71 / 72 | 12 |
| Kimi K3.1 Pro | 70.5 / 72 | 12 |
| Qwen3.8 Max 0902 | 69 / 72 | 12 |
| Gemini 4 Argon | 66 / 72 | 12 |
| GLM-5.3 | 62.5 / 72 | 12 |
API test dated 24 October 2026: 12 tasks per language, one attempt, maximum 72 points. Standard Kimi endpoint: moonshotai/mxfp4. Highspeed was tested only in Polish. Differences of 1-2 points are uncertain; this is not a web app test. A Claude Opus agent prepared the tasks; automatic checks and AI judges scored them.
Comparison scheduled: 2026-11-03 09:30:00