Kimi K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, fourth in its launch table and 3 points behind Claude Opus 5.5. Moonshot AI released it and a faster Highspeed variant on 23 October 2026 and promises the full weights by 4 November 2026, but they were not published at the time of writing.
The short answer
- Released: 23 October 2026 (Artificial Analysis and Moonshot’s launch post).
- Independent score: 55 on the Artificial Analysis Intelligence Index, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56), and above Kimi K3 (44).
- Context: 1M tokens.
- API price per 1M tokens: Pro costs $2 input (cache miss), $0.20 cache hit and $10 output. Highspeed costs $4, $0.20 and $20.
- Open weights: promised “by Nov 4, 2026”, not published at the time of writing, and no license named.
- Size (Artificial Analysis): 2.8T total parameters, 104B active; text and image input.
This guide separates Moonshot’s own claims from the independent results, as read on 25 October 2026. The model was not run.
What does the independent evaluation show?
| Model (as named by Artificial Analysis) | Intelligence Index |
|---|---|
| Claude Opus 5.5 (max with fallback) | 58 |
| Claude Sonnet 5.5 (max with fallback) | 56 |
| Muse Wonder Full (max) | 56 |
| Kimi K3.1 Pro | 55 |
| Kimi K3 (max) | 44 |
The jump from Kimi K3 is 11 points (55 minus 44, our subtraction). The gap to the leader is small, 3 points. So it sits in the same band as the top models on this index, not clearly above or below.
Artificial Analysis also reports $1.98 per Intelligence Index task, roughly equal to Kimi K3. K3.1 Pro uses more output tokens (55K per task against 48K), even though its per-token prices are 33% lower. A lower token price therefore did not mean a lower cost per task. That is the point of why a cheaper AI model can cost more per finished task.
These results come from one dated launch snapshot. They are not a general accuracy rate and not a test of Arabic. For how to read tables like this, see five ways AI benchmark tables mislead.
Why are there two Terminal-Bench numbers?
Artificial Analysis reports 67% for K3.1 Pro on Terminal-Bench 4.0, ahead of Claude Sonnet 5.5 at 64%, and calls it the highest score of any model it measured there. Moonshot’s own post reports 65.7%, “ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%)”, run with the Claude Code harness.
The figures come from different setups, so do not merge them into one ranking. Moonshot also warns that testing of proprietary models may involve some fallback behavior.
How fast is it, and where can you use it?
Moonshot says Highspeed “delivers 2-3x the speed of K3.1 Pro”. That is a company claim with no independent measurement found. Highspeed costs double the Pro price per token ($4 and $20 against $2 and $10), so you pay for the speed.
Moonshot lists Kimi.ai and the mobile apps, the Kimi Work desktop app (3.1.0 or later), Kimi Code (the /model command) and the API. The model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Zero Data Retention is available for API customers. Kimi Enterprise administrators must enable the model because access is off by default at launch.
For price comparison with other APIs, see DeepSeek API models and pricing and Claude plans and pricing.
What else does Moonshot report?
These results are company-reported, not independent.
- Cyber capability: Moonshot says it removed cyber content from training for K3.1 Pro. On its internal xdev-bench-hard it reports 81.0% pass@3, and says a version without that data removal scored 90.5%. It says the model’s cyber capability remains behind frontier proprietary models.
- Hallucination: on its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% (K3) to 8.3%. The company notes this is a stress test, not a typical daily error rate.
- CAD: BenchCAD (with tools) 92.7%, against Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9% in Moonshot’s table.
Several of these benchmarks are internal or run with Moonshot’s own harnesses, so treat them as claims until independent results appear.
Who should wait, and who can try it now?
If you need open weights to host the model yourself, wait: they are promised, not published, and no license is named. If you only need an API, you can test it now.
- Open weights and license: Moonshot promises the weights by 4 November 2026 and names no license. Check its page before building on them.
- Arabic quality and your own projects: the evaluations read do not cover them. We suggest testing a few of your own tasks, as described in how to run a one-afternoon AI model evaluation.
Questions and answers about Kimi K3.1 Pro
Is Kimi K3.1 Pro open-weight today?
No. Moonshot promises to release the full weights by 4 November 2026. They were not published at the time of writing.
Is it better than Claude Opus 5.5?
Not on this index. It scored 55 against 58 for Claude Opus 5.5 in the same launch table.
How much does the API cost?
For Pro, $2 per million input tokens (cache miss), $0.20 cache hit and $10 per million output tokens, according to Kimi’s pricing page. Highspeed is $4, $0.20 and $20.
Is Highspeed worth the price?
It costs twice as much per token for a claimed 2-3x speed. Pay for it only if latency is your bottleneck.
Terms in plain language
- Open weights: the trained model files are published so others can download them.
- Cache hit: input the service has already processed and bills at a lower price.
- Harness: the tools and rules around a model during a test.
- Pass@3: a score that counts a task as solved if any of three tries succeeds.
Sources
- Artificial Analysis: Kimi K3.1 Pro launch benchmarks (23 October 2026)
- Moonshot AI: Kimi K3.1 Pro
- Kimi API pricing
Facts read on 25 October 2026.
Part of the guide: AI Model Pricing and API Cost Guides
