Moonshot AI’s Kimi K3.1 Pro has landed fourth on the Artificial Analysis Intelligence Index with 55 points, just 3 points behind the leading proprietary models, according to The Decoder’s analysis of the independent benchmark published October 23. The Chinese lab’s new 3T-class model jumped 11 points over its Kimi K3 predecessor and trails only Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full. Moonshot also revealed it deliberately stripped cyber-related data from both training stages of the open-weights release, cutting the model’s vulnerability-research score from 90.5% to 81.0% on its internal xdev-bench-hard test. If the weights arrive by November 4 as planned, K3.1 Pro would become the highest-scoring open-weights model available.
The Intelligence Index from Artificial Analysis aggregates ten evaluations — from agentic coding and terminal use to long-context reasoning and knowledge work — into a single score. Kimi K3.1 Pro’s 55 puts it three points behind Claude Opus 5.5 at 58 and one point behind Claude Sonnet 5.5 and Muse Wonder Full at 56, and, according to The Decoder, no model with a comparable parameter count scores higher. On the GDPval-AA agentic benchmark the model reached 1,846 Elo, 312 points above K3 and just behind Opus 5.5’s 1,866.
Agentic Tasks See The Biggest Gains
Moonshot’s own numbers, reported with reasoning effort set to maximum, show the sharpest improvements where agents do real work. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra’s 58.2% and narrowly behind Claude Opus 5.5’s 66.4% — a remarkable leap from K3’s 13%. On BenchCAD, which tests 3D reconstruction through CAD code generation, it hits 92.7% with tools, approaching Opus 5.5’s 96.2%. Hallucinations fell too: claim-level errors on the company’s internal Kimi Hallu Bench Hard dropped from 23.4% with K3 to 8.3%.
The model keeps K3’s dimensions — 2.8 trillion total parameters with 104 billion active, and a 1-million-token context window — while token prices fall by a third: input costs $2 per million tokens, output $10, and cache hits $0.20. A companion variant, K3.1 Pro Highspeed, runs two to three times faster for more responsive interactions.
Cyber Data Deliberately Stripped
The most unusual part of Moonshot’s technical blog is an entire section on cybersecurity. The company says it watched vulnerability-research capability grow steadily as post-training scaled, traced the trend to cyber-related content embedded in the general mid-training corpus and cyber tasks in the reinforcement-learning mix, and decided to remove that data from both stages for the open-weights release.
The effect is measurable: on xdev-bench-hard, an in-house benchmark that reproduces recently disclosed vulnerabilities in software like the Linux kernel and Node.js, the uncleaned checkpoint solved 90.5% of tasks while the final K3.1 Pro manages 81.0%. Moonshot’s stated principle is that no open-weights model should push the absolute frontier of offensive cyber capability — while still raising what open models can do for defensive work like auditing legacy code.
Why It Matters
Kimi K3.1 Pro is doing for open weights what DeepSeek did for efficient training: compressing the gap between what anyone can download and what only the biggest labs can run. Shrinking the open-proprietary gap from 12 points to 3 on an independent benchmark is the kind of result that reshapes procurement decisions, especially at a third of the previous token price — and it builds directly on the K3.1 Pro launch covered here yesterday.
The deliberate cyber-data removal is arguably the more consequential move. It is one of the first times a frontier lab has publicly quantified exactly how much capability it is choosing not to ship — and put a number on the trade-off. As regulators from Beijing to Brussels weigh rules on open models, Moonshot has handed them a worked example of self-restraint with the receipts attached. For the open-weight versus closed-model debate, it reframes the question from whether open models can compete to what their makers are willing to withhold.
Background
Moonshot AI launched Kimi K3.1 Pro on October 23 as a post-trained version of Kimi K3 with substantially larger post-training compute, available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full weights promised by November 4. The release follows a year in which Chinese open-weight models — from DeepSeek to Qwen to MiniMax — have steadily closed the distance to Western proprietary systems on independent benchmarks.
Source: The Decoder ; Moonshot AI
0 Comments