Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, three points below Claude Opus 5.5, at half the price per token. Moonshot AI put it on its API on 23 October 2026 and promises the full weights by 4 November. If that happens, it becomes the highest-scoring open-weights model on the index, and the gap between open and closed models shrinks from 12 points to 3.
What Kimi K3.1 Pro is
This isn't a new architecture. K3.1 Pro is Kimi K3 with far more post-training compute: the same 2.8 trillion total parameters and 104 billion active per token, a one-million-token context window, and text plus image input. The difference is what it does with them: 11 points more than K3 on the Artificial Analysis index, good for fourth place behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56).
The biggest gains are in agentic work. On Terminal-Bench 4.0, Moonshot reports 65.7%, up from K3's 13%, ahead of GPT-6 Astra (58.2%) and just short of Opus 5.5 (66.4%). On GDPval-AA, as measured by Artificial Analysis, it reaches 1,846 Elo, 312 points above K3 and 20 behind Opus 5.5 (1,866).
Hallucinations drop too on Moonshot's internal Kimi Hallu Bench Hard: from 23.4% of false claims with K3 to 8.3%. GPT-6 Astra sits at 5.5% on the same test, so this is not where Kimi leads.
There's also a K3.1 Pro Highspeed variant which, according to Moonshot, responds two to three times faster and costs twice as much, except for cache hits, which stay at the same price.
How to try it
Moonshot's API is OpenAI-compatible and the model id is kimi-k3.1-pro. The quickest way to send it a prompt from a terminal is Simon Willison's llm tool (version 0.32 or later), which has a command for any compatible endpoint:
llm keys set kimi # paste your Moonshot API key
llm openai endpoint \
https://api.moonshot.ai/v1 \
--key kimi \
-m kimi-k3.1-pro \
--responses \
-o reasoning_effort medium \
"Explain what this cron schedule does: */15 2-6 * * 1-5"The model supports three reasoning levels: low, medium and max. Moonshot's benchmark numbers use max, which also produces the most output tokens; if you're comparing cost against your current model, measure at the level you'll actually run.
Kimi K3.1 Pro vs Claude
Anthropic is the obvious comparison, since both Moonshot and Artificial Analysis place it next to Opus 5.5. But the more telling one is Sonnet 5.5:
| Kimi K3.1 Pro | Claude Opus 5.5 | Claude Sonnet 5.5 | |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 55 | 58 | 56 |
| Input (per million tokens) | $2 | $4 | $2 |
| Output (per million tokens) | $10 | $20 | $10 |
| Cache read (per million) | $0.20 | $0.20 | $0.10 |
| Terminal-Bench 4.0 | 65.7% | 66.4% | — |
| Weights | Promised by 4 November | Closed | Closed |
At these prices, a service that burns 10 million input tokens and 2 million output tokens a day (no caching) would pay $40 a day with Kimi K3.1 Pro or Sonnet 5.5, and $80 with Opus 5.5. Highspeed, at double the rate, lands at $4 and $20: exactly Opus 5.5 pricing.
Moonshot also cut prices by 33% compared with K3 (input from $3 to $2, output from $15 to $10), though cost per task barely moves: Artificial Analysis puts it at $1.98 versus $2.00, because K3.1 Pro writes 14% more output tokens.
Where it wins and where it doesn't
It wins on score per dollar against Opus 5.5. Three points lower on the index for half the cost is a fair deal for long agentic tasks, where output tokens drive the bill. On Terminal-Bench 4.0 the gap to Opus 5.5 is seven tenths of a point.
It doesn't beat Sonnet 5.5 on price. They cost the same per token, Sonnet is a point higher on the index and its cache reads are half the price. If you'll only ever use it through an API, the deciding factor is how each one handles your workload, not the price sheet.
Its real edge is the weights. A frontier-grade model you can deploy on your own infrastructure, fine-tune, or buy from whichever provider you like changes your negotiating position with every closed API. That said, 2.8 trillion parameters won't fit on a lab box: open weights matter mostly to inference providers and companies with their own GPU clusters, and for everyone else they'll mean more providers serving the model.
The open version won't be identical on security work. Moonshot says that, to release the weights, it removed cybersecurity data from both training phases. On its internal benchmark for reproducing recently disclosed vulnerabilities (xdev-bench-hard), the unfiltered version solves 90.5% of tasks and the filtered one 81.0%. If your use case is code auditing or vulnerability analysis, test the exact model you plan to run.
Verdict
If you already run Opus 5.5 on agentic tasks and cost matters, K3.1 Pro deserves a run against your own task suite: the benchmark gap is small and the bill is half. If you're on Sonnet 5.5, switching for API use alone has no clear economic case. And if you need a model you can host or fine-tune, the date to watch is 4 November 2026: if the weights ship with this score, it'll be the best open option available.
Sources
- Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5, The Decoder, 23 October 2026.
- Introducing Kimi K3.1 Pro, Simon Willison, 23 October 2026.
- Claude pricing, Anthropic documentation.