Kimi K3.1 Pro scores 55 on Artificial Analysis and adds a 2-3x Highspeed tier

Artificial Analysis scores K3.1 Pro 55 on its Intelligence Index, 4th of 215 models, and lists it at $2.00 per million input tokens and $10.00 per million output tokens. For coding agents, Kimi reports 65.7% on Terminal-Bench 4.0. That is behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%, which Kimi describes as matching or approaching the strongest proprietary models.

Kimi calls it the first model in its K3.1 series and its most capable 3T-class open-weight model. It is live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. On the API it is selected as kimi-k3.1-pro or kimi-k3.1-pro-highspeed, with Zero Data Retention available to API customers. Kimi Enterprise administrators have to switch it on, because access is off by default at launch.

Kimi describes K3.1 Pro as a post-trained version of Kimi K3 with substantially larger post-training compute. To run large-scale joint RL across hundreds of downstream domains, Kimi says it optimized every layer of the stack: the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks. The stated aim was that each increment of RL compute delivers stable, predictable returns. Artificial Analysis lists text and image input, text output and a 1M-token context window.

The sources do not name a licence for the weights. Artificial Analysis currently labels the model proprietary and does not say why. Its legend separates open weights from 'Commercial Use Restricted' and 'Non-commercial' licences. Teams planning to self-host should wait for the November 4 release and read the licence terms before building around it.

The title of the Artificial Analysis source says the model is placed on the Pareto line. On the page, the Pareto line appears on the Intelligence Index vs. Cost per Task chart. The index, v4.3.2, combines 10 evaluations.

Artificial Analysis labels four of those evaluations agentic. AA-Briefcase v1.1 covers agentic knowledge work, GDPval-AA v2.1 covers agentic real-world work tasks, AutomationBench-AA covers agentic SaaS workflows, and Terminal-Bench 4.0 covers agentic coding and terminal use. The other six are SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1.

The score is for the reasoning version of the model. Artificial Analysis compares proprietary models against others in the same price range, using a blended 3:1 input/output price ratio.

One commenter on Hacker News wrote that the model probably does not move the frontier and instead sits at the current frontier, which they put at Opus 5.5. The page text in the source does not show Artificial Analysis's per-evaluation scores. That includes its own Terminal-Bench 4.0 result, so the source does not show which of the 10 evaluations contribute most to the 55.

Kimi ran Terminal-Bench 4.0, DeepSWE v1.1 and SWE-Marathon v1.1 with the Claude-Code harness. Unless otherwise specified, all K3.1 Pro results use reasoning effort set to max, temperature 1.0 and top-p 0.95.

On Terminal-Bench 4.0, most competitor scores are the maximum value reported in either the model's blog post or the official leaderboard. Kimi evaluated DeepSeek-V4.1-Flash-1017 itself on SWE-Marathon and PostTrainBench. PostTrainBench was run on H20 GPUs instead of H100s, with DeepSeek-V4.1-Flash-1017 at maximum effort as the judge. Kimi says it chose that judge because it best matched the official GPT-5.6-Terra (xhigh) judge in testing.

A team running K3.1 Pro at lower effort or through a different harness should not assume these numbers carry over. Artificial Analysis does not state, in the source, which effort setting produced its 150M-token verbosity figure.

For 3D work, K3.1 Pro scores 92.7% on BenchCAD with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.

Kimi also describes a long-horizon internal test: building a human walking dynamics simulator on top of the open-source human-atlas model. Across 167 missions and 420 independent reviews, the model took the simulated human from collapsing within seconds to walking for 11.2 seconds. Along the way it identified and corrected 36 errors in shoulder-muscle moment-arm calculations.

Kimi's only quantitative claim for Highspeed is that it delivers 2-3x the speed of K3.1 Pro. Every benchmark in the blog is reported for K3.1 Pro, and the Artificial Analysis page covers K3.1 Pro only. Neither source gives Highspeed's price, benchmark scores, or how the speed is achieved. Applied to Artificial Analysis's 73.5 tokens per second, the multiplier would mean roughly 147 to 220 tokens per second, but no source measures it.

K3.1 Pro produced 150M output tokens across Artificial Analysis's index and runs below average on speed, so throughput is a real constraint for long agent loops. Kimi attributes K3.1 Pro's agentic gains to RL post-training, and nothing published shows whether Highspeed matches K3.1 Pro on quality. Until scores appear, run your own task suite against both kimi-k3.1-pro and kimi-k3.1-pro-highspeed before routing work to the faster tier.

For the open-weight release, Kimi identified and removed cyber-related data from both the mid-training corpus and the RL environment mix, while keeping general-purpose coding tasks. Kimi tested the result on xdev-bench-hard, its in-house exploit-development benchmark. Most of its tasks reproduce recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL, with mitigations enabled.

A checkpoint trained without these measures solves 90.5% of problems at pass@3, and the released K3.1 Pro solves 81.0%. Kimi says this is still a substantial improvement over K3.1 Pro's predecessor. It also says the model remains behind frontier proprietary models that have been publicly available for at least three months.

One commenter on Hacker News called the public version "nerfed" in cybersecurity like the US models. Another said Chinese models had been attractive for security-adjacent work and red-teaming because they did not trip their own guardrails. The method Kimi describes is data removal at two training stages, and the blog does not discuss refusal training. A third commenter drew that distinction, describing it as downweighting offensive cyber as a post-training domain rather than active nerfing.

Another commenter asked how much more performance could be gained by further RL post-training on the open weights. The sources do not address this, and it becomes a practical question once the weights ship.

For defensive teams, the published figure is 81.0% pass@3 on xdev-bench-hard. That is reported over a random subset of runs, with no hard timeouts or output token limits. Kimi has not published how the model responds to refusal-style security prompts, so security-adjacent use cases need their own evaluation.