Chinese AI lab Moonshot AI released Kimi K3.1 Pro on Thursday, the first model in its new K3.1 series and what the company calls its most capable open-weight model to date, with a 1-million-token context window. The launch also introduces Kimi K3.1 Pro Highspeed, a variant the company says runs two to three times faster than the standard model. Both are available immediately on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with the full model weights scheduled for release by November 4.
The release lands at a charged moment in the open-model race, days after DeepSeek's reported $15 billion funding round tightened the competition among China's frontier labs. Bloomberg framed the launch as Moonshot's first major upgrade since its own "DeepSeek moment," and the model drew immediate attention on Hacker News, where it collected nearly 400 points within hours.
Fourth Place on the Artificial Analysis Index
The headline result comes from the Artificial Analysis Intelligence Index, where Kimi K3.1 Pro scores 55 points — an 11-point jump over its predecessor, Kimi K3, that places it fourth overall. Only three proprietary models rank higher: Anthropic's Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56, according to Artificial Analysis.
The gap matters most for what happens next. If Moonshot ships the weights on schedule by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model available, shrinking the distance between the best open and proprietary models from 12 points to 3 — assuming nothing else changes in the intervening weeks.
The Decoder reported that the parameter count stays at 2.8 trillion, while Moonshot describes the system as a post-trained version of Kimi K3 built with substantially larger post-training compute. The company says it ran large-scale joint reinforcement learning across hundreds of downstream domains, reworking the RL framework, long-context inference and sandbox infrastructure so that each increment of compute delivered stable returns.
Agentic Coding Is the Star of the Show
K3.1 Pro's strongest gains show up in agentic tasks. On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks, the model scores 65.7% — ahead of OpenAI's GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%. For perspective, Kimi K3 managed just 13% on the same test.
On Moonshot's internal Kimi Code Bench 3.0 Long-horizon, which measures multi-step execution and end-to-end problem solving across engineering domains, K3.1 Pro scores 61.2% against K3's 55.95%. On the GDPval-AA agentic benchmark tracked by Artificial Analysis, the model reaches 1,846 Elo — 312 points above K3 and within reach of Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857.
Hallucination rates dropped sharply as well. On the company's internal Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4% with K3 to 8.3% — though Moonshot cautions the benchmark is a deliberate stress test rather than a measure of general error rates, and GPT-6 Astra still leads at 5.5%.
3D Scenes, Game Engines and Financial Research
Beyond raw benchmarks, Moonshot is positioning K3.1 Pro as a creative and professional tool. The company demonstrated the model building complete experiences across game engines including Unity, Unreal Engine and Godot, combining them with web search and 3D generation. Its front-end output — interactive Three.js scenes and 3D assets — shows stronger composition, typography and animation than its predecessor, according to the company's tech blog.
Moonshot also reports gains on internal finance benchmarks, with K3.1 Pro improving 7.2 points on Financial Modeling and 11.6 points on Financial Research over K3, and topping all tested models on the OfficeQA Pro v2 office-tasks benchmark. As with all vendor-reported internal evaluations, independent verification will be the real test.
Pricing Dropped 33%
Alongside the capability jump, Moonshot cut prices. K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 — a roughly 33% reduction from K3's API pricing, according to The Decoder. The Highspeed variant runs $4 per million input tokens and $20 per million output tokens in exchange for its two-to-three-times faster inference.
That pricing puts K3.1 Pro well below the flagship tiers of the proprietary models it now trails on leaderboards, a combination that explains much of the developer community's enthusiasm. For teams running agentic workloads that burn through millions of tokens, the difference between $10 and $25 per million output tokens compounds quickly.
Why the November Weights Release Matters
The open-weights release planned for November 4 is arguably the story's real pivot point. Kimi K3's weights appeared on Hugging Face earlier this year under a permissive license and were quickly adopted by enterprises running self-hosted deployments. If K3.1 Pro ships as promised with its fourth-place score intact, it would hand developers a downloadable model within striking distance of Claude Opus 5.5 — a milestone no open release has reached before.
Rivals are not standing still. DeepSeek is reportedly raising at more than double its original target, Qwen's latest series continues to post strong open-weight numbers, and Western labs keep moving the proprietary ceiling. But the trajectory is clear: the gap that measured 12 points on the Artificial Analysis Index this week may be measured in single digits — or smaller — by year's end.
For now, K3.1 Pro and K3.1 Pro Highspeed are live on Kimi's consumer products and API, with the weights countdown to November 4 already running.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
