Moonshot AI releases Kimi K3.1 Pro open-weight model with 1M-token context

Chinese AI lab Moonshot AI launched Kimi K3.1 Pro on October 23, its most capable open-weight model to date, featuring a 1-million-token context window, a mixture-of-experts architecture with 2.8 trillion total parameters (104 billion active per token), and pricing that undercuts proprietary rivals. The company promises full model weights by November 4. Moonshot's own benchmarks claim the model matches or beats Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on several hard tests. Independent evaluations show a strong but not frontier-leading model, with notable gaps in scientific reasoning and a pricing structure whose real-world cost depends heavily on how much the model "thinks."

What shipped and when

Kimi K3.1 Pro and a faster "Highspeed" variant went live on October 23–24 across Kimi.ai, Kimi Work, Kimi Code, and the Kimi API [7][8]. The model is a post-trained version of Kimi K3 (released July 16), keeping the same 2.8T/104B MoE architecture and 1M-token context but trained with "substantially larger post-training compute" across hundreds of downstream domains [7][3]. Moonshot says the full weights will be released by November 4 under a license not yet disclosed [7][3]. Enterprise access is off by default at launch [7].

API pricing: $2 per million input tokens and $10 per million output for the standard model; $4/$20 for Highspeed (which Moonshot says runs 2–3× faster). Cache hits cost $0.20 per million tokens. Access requires a minimum $1 top-up [8][4]. Temperature is fixed at 1.0, top-p at 0.95, and reasoning effort defaults to "max" [8].

Moonshot's benchmark claims

Moonshot's launch blog and technical report compare K3.1 Pro against Opus 5.5, GPT-6 Astra, Claude Fable 5.1, DeepSeek-V4.1-Flash, and GLM-5.3 on 14 benchmarks, mostly run by Moonshot itself on the Claude Code harness at maximum reasoning effort [7][2]. Highlights from Moonshot's table:

• Terminal-Bench 4.0 (coding agents): 65.7%, behind Opus 5.5 (66.4%) but ahead of GPT-6 Astra (58.2%) [7][2]
• SWE Marathon v1.1 (long-horizon software engineering): 51.7%, leading all models in the table [7][2]
• Humanity's Last Exam with tools: 70.0%, ahead of Opus 5.5 (67.7%) and GPT-6 Astra (57.2%) [7][2]
• BenchCAD with tools (3D reconstruction via CAD code): 92.7%, approaching Opus 5.5 (96.2%) and Astra (95.9%) [7][2]
• OfficeQA Pro v2 (financial document analysis): 64.4%, reported as state-of-the-art [7][3]
• Hallucination stress test (Kimi Hallu Bench Hard): claim-level hallucinations fell from 23.4% (K3) to 8.3% [7][3]

Moonshot also reports large gains on internal finance benchmarks (+7.2 points financial modeling, +11.6 points financial research) and a 4.2× improvement on a simulated business-operation benchmark [7][3].

Caveats: Nearly all these figures come from tests Moonshot ran itself. Several benchmarks (Kimi Code Bench, Kimi Hallu Bench, OfficeQA Pro, xdev-bench-hard) are internal. Rival scores mix Moonshot's own runs, vendor system cards, blogs, and public leaderboards. Moonshot notes that "proprietary model testing may involve some fallback behavior" and that GPT results "may include cyber safeguards" [7].

Independent evaluations: strong, not dominant

Three independent sources paint a more modest picture.

Artificial Analysis (October 23) scores K3.1 Pro at 55 on its Intelligence Index — an 11-point jump over K3 — placing it fourth among 215 models, behind Opus 5.5 (58), Sonnet 5.5 (56), and Muse Wonder Full (56) [2][6]. If weights ship as planned, it would become the highest-scoring open-weight model, narrowing the open-vs-closed gap from 12 to ~3 points [2]. Artificial Analysis measured 150M output tokens to complete its index (median: 87M), 77 tokens/sec output speed, and a weighted average cost of $2.11 per index task vs. $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra [6].

Arena's web development leaderboard (blind human preference votes) ranks K3.1 Pro eighth, behind Opus 5.5, GPT-6 Astra, Sonnet 5.5, GPT-6.1 Soul, Claude Fable 5.1, Muse Wonder Full, and another Sonnet variant — up from 17th for K3 [1].

KingBench (a hands-on coding benchmark of six practical tasks: retro emulator, container runtime, 3D room designer, keyboard config app, Pokémon clone, fitness tracker) scored K3.1 Pro 53.7% (5.37/10), sixth of eleven models tested. It trailed Opus 5.5 (5.78), GPT-6 Astra (5.65), GPT-6.1 Soul (5.50), Gemini 4 Argon (5.43), and Claude Haiku 5.5 (5.42) [1]. Moonshot's blog had claimed a seven-point lead over GPT-6 Astra; the independent test showed Astra clearly ahead.

Where it falls short: science and scope completion

The weakest independent result is Terminal-Bench Science 0.1 (scientific research workflows in a terminal): K3.1 Pro scores 38.1% vs. 68.1% for GPT-6 Astra, 63.3% for Opus 5.5, and 40.0% for Fable 5.1 [7][6]. Moonshot's own table shows the same gap [7].

KingBench revealed a consistent pattern: K3.1 Pro builds solid foundations but often fails to complete the full scope. Asked to build Game Boy, DS, and 3DS emulators in Rust, it delivered only the Game Boy core (though it was honest about the omission, and the core passed classic CPU instruction tests). A "Docker alternative" pulled and verified container images — even blocking 20 malicious ones — but had no build, exec, or networking functionality despite the readme claiming otherwise [1]. The model's best result was a keyboard-config app that correctly implemented the VIA/QMK hardware protocol most models get wrong.

This honesty about limitations aligns with Moonshot's claimed hallucination improvements: the model avoids confidently faking functionality it didn't build [1].

Cybersecurity: deliberate capability capping

Moonshot devotes a full section of its technical blog to cyber risk. It found that vulnerability-research performance improved steadily as post-training scaled, traced to cyber-related content in the mid-training corpus and RL environment mix. For the open-weight release, Moonshot stripped cyber data from both stages [7][3].

On its internal xdev-bench-hard (reproducing recently disclosed vulnerabilities in Linux kernel, Node.js, JavaScriptCore, etc.), an unfiltered checkpoint solved 90.5% at pass@3; the released K3.1 Pro solves 81.0%. DeepSeek-V4.1-Flash reaches 85.7% on the same test; K3 scored 33.3% [2]. Moonshot states K3.1 Pro's offensive cyber capability remains behind frontier proprietary models public for at least three months, and frames gradual diffusion as better for defensive security than concentration in closed systems [7].

Pricing: cheap tokens, expensive thinking

Headline API prices are roughly half of Opus 5.5 ($4/$20) and a fifth of reported GPT-6 Astra pricing ($10/$50) [1][6]. But KingBench found K3.1 Pro is a "chatty" agent that rereads large cached context on every step. Across six tasks using ~112M tokens, the run cost ~$32, with most cost from rereading cache rather than generating output. Compared to Claude Haiku 5.5 (similar price, slightly higher KingBench score, ~¼ the cost) and DeepSeek V4.1 Flash (close score, under $2), K3.1 Pro wasn't the economical choice in pure API terms despite low per-token rates [1].

The pricing case improves on a flat-rate Kimi subscription (where per-token costs don't apply) or once open weights enable self-hosting — though the 2.8T scale points to multi-node "supernode" serving infrastructure [4][5].

What's unknown

• The open-weight license (K3 used the "Kimi K3 licence"; Moonshot hasn't published K3.1 Pro's) [5].
• Whether the downloadable weights will exactly match the hosted API version (Moonshot says it removed cyber-related training data for the open release) [3].
• Reproduction of Moonshot's benchmark numbers by third parties (none have yet) [5].
• Real-world performance on the "hundreds of downstream domains" Moonshot cites beyond the tested benchmarks [7].

Bottom line

Kimi K3.1 Pro is a genuine advance: a 2.8T MoE model with 1M context that delivers strong agentic coding, 3D/CAD, and finance performance at disruptive pricing, with open weights promised in weeks. Independent tests confirm it's competitive — fourth on Artificial Analysis, eighth on Arena, sixth on KingBench — but not the frontier-beater Moonshot's selected benchmarks suggest. The science-reasoning gap (Terminal-Bench Science 38% vs. 68% for Astra) and the "chatty" token economics are real limitations. If the November 4 weight release happens as promised, it will be the highest-scoring open-weight model on at least one major index, narrowing the open/closed gap significantly. That matters for organizations that need near-frontier capability on their own infrastructure.