Image: Moonshot AI — Kimi K3.1 Pro (official product image). Moonshot's official benchmark overview: Kimi K3.1 Pro (blue) against Claude Opus 5.5, GPT-6 Astra and others across 14 evaluations.
TL;DR — Kimi K3.1 Pro release benchmarks, explained: Moonshot AI's new open-weight model launched on 23 October 2026, scoring 65.7% on Terminal-Bench 4.0 — ahead of GPT-6 Astra (58.2%) and 0.7 points behind Claude Opus 5.5 (66.4%). With open weights promised by 4 November and API pricing 33% below Kimi K3, it is the closest an open model has come to the closed frontier.
📌 KEY HIGHLIGHTS
- Moonshot AI launched Kimi K3.1 Pro on 23 October 2026 — a post-trained upgrade of Kimi K3, not a new base model (Moonshot).
- Artificial Analysis ranks it 4th on its Intelligence Index at 55 — behind only Claude Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56); once weights ship it becomes the top open-weights model.
- Specs: 2.8T total / 104B active parameters (MoE), 1M-token context, text + image inputs; a Highspeed variant runs 2–3× faster.
- API pricing cut by a third: $2 per 1M input / $10 per 1M output tokens; full weights free to download by 4 November 2026.
- Caveat: most benchmark numbers are Moonshot's own and not yet independently verified — science tasks remain a weak spot.
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is the first model in Moonshot AI's new K3.1 series — the Beijing lab's "most capable 3T-class open-weight model to date". It is not a from-scratch model: Moonshot describes it as a post-trained version of Kimi K3 (released July 2026), pushed through far more reinforcement learning across what the company calls "hundreds of downstream domains". The architecture is a mixture-of-experts design with roughly 2.8 trillion total parameters, 104 billion active per token, a 1-million-token context window, and multimodal text+image input.
It is live now across the Kimi family — the Kimi app and website, the Kimi Work desktop app (v3.1.0+), the Kimi Code coding tool and the Kimi API (identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed). Moonshot also shipped 3D Make on kimi.ai — preview and edit AI-generated 3D models online, then download them for 3D printing. For more on the open-model wave, see our NVIDIA's open models for AI agents explainer.
Kimi K3.1 Pro release benchmarks: the numbers
The headline result is agentic coding. On Terminal-Bench 4.0 — coding agents doing realistic terminal tasks — K3.1 Pro scores 65.7%, beating OpenAI's GPT-6 Astra (58.2%) and landing 0.7 points behind Claude Opus 5.5 (66.4%). Independent evaluator Artificial Analysis confirms the trajectory: its Intelligence Index score jumped 11 points over K3 (44 → 55), with the biggest gains in agentic tests.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
| Terminal-Bench 4.0 (coding agents) | 65.7% | 66.4% | 58.2% |
| Artificial Analysis Intelligence Index | 55 (4th) | 58 | 53 |
| GDPval-AA (work tasks, Elo) | 1,846 | 1,866 | 1,573 |
| BenchCAD with tools (3D/CAD) | 92.7% | 96.2% | 95.9% |
| OfficeQA Pro v2 (finance docs) | 64.4 (best reported) | 62.2 | 54.4 |
| Terminal-Bench Science 0.1 | 38.1% | 63.3% | 68.1% |
Image: Moonshot AI — Kimi K3.1 Pro (official product image). BenchCAD results with tools: Kimi K3.1 Pro at 92.7%, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
Two more numbers matter. On Moonshot's internal hallucination stress test, the claim-level error rate fell from 23.4% (K3) to 8.3% — a ~65% drop. And on the cyber front, Moonshot says it deliberately stripped cyberattack-related data from training so the open release won't advance offensive hacking: the shipped model scores 81.0% on its internal xdev-bench-hard versus 90.5% for the unfiltered checkpoint.
Pricing and where to try it
Moonshot cut API prices by a third versus K3. The standard tier costs $2 per million input tokens, $10 per million output, $0.20 cached input; the Highspeed tier doubles those rates for 2–3× faster responses. That undercuts typical frontier-model pricing while matching much of the capability — the pitch Indian startups and indie developers will notice, especially with Indian tech press already covering the launch.
| Tier | Input / 1M tokens | Output / 1M tokens | Speed |
| Kimi K3.1 Pro | $2.00 | $10.00 | Standard |
| Kimi K3.1 Pro Highspeed | $4.00 | $20.00 | 2–3× faster |
For now the only way to run it is Moonshot's hosted API and apps. The full open weights land by 4 November 2026 — at which point anyone with enough hardware can self-host, and Artificial Analysis says it will become the highest-scoring open-weights model, narrowing the open-vs-closed gap from 12 points to 3.
The honest caveats
Three things to keep in perspective. First, nearly all benchmark figures are Moonshot's own — measured at maximum reasoning effort, with rivals' scores taken from their published reports. Independent verification will come once the weights ship and the community can run its own tests. Second, science is a clear weak spot: 38.1% on Terminal-Bench Science 0.1 against 68.1% for GPT-6 Astra. Third, the cyber-data removal means the open model is deliberately held below what Moonshot's pipeline can produce — a safety choice, but also a capability ceiling worth knowing about.
Update (24 October): Reuters confirms the curb — and the debate it feeds
Reuters reported on 24 October 2026 that Moonshot AI has confirmed it removed cyber-related data from Kimi K3.1 Pro's mid-training and reinforcement-learning environments, while keeping its general coding skills. The company's stated principle: "we aim not to release open weights that would extend the absolute frontier of cyber capability." The shipped model scores 81.0% on Moonshot's internal xdev-bench-hard exploit test, against 90.5% for the unfiltered checkpoint.
The disclosure lands days after Beijing proposed draft export controls covering AI-driven malware-generation and mutation technology and restrictions on AI-assisted vulnerability-discovery tools — measures still open for public comment, citing risks to critical infrastructure, defence-industrial and core government systems.
It also feeds an open debate. Anthropic and OpenAI have argued in recent weeks that downloadable open models raise cyber risk, since anyone can modify them; Moonshot's counter is that open weights let more teams audit and secure their own software, including legacy systems that receive little attention. One caveat the disclosure does not address: the controls apply at training time, so they do not by themselves stop someone fine-tuning the released weights to restore the filtered capability. Moonshot says it intends to raise the cyber capability in its open weights over time.
Should you care? 🤔
🤔 YES — if you build with AI. Kimi K3.1 Pro is the first open-weights model within touching distance of the closed frontier on agentic coding, at a third less than its predecessor's API price — and the weights go free on 4 November. For developers and startups (including India's), that means frontier-class coding agents you can eventually self-host. Just don't treat Moonshot's self-reported scores as final until independent testers reproduce them — and note the science gap if your workload is research-heavy. Compare with Claude running on Google Cloud for the enterprise-managed alternative.
FAQs
When was Kimi K3.1 Pro released?
Moonshot AI released Kimi K3.1 Pro on 23 October 2026 across the Kimi app, Kimi Work, Kimi Code and the Kimi API. The full open model weights are promised by 4 November 2026.
How does Kimi K3.1 Pro compare to Claude Opus 5.5?
On Terminal-Bench 4.0, K3.1 Pro scores 65.7% vs 66.4% for Opus 5.5 — just 0.7 points behind — and beats GPT-6 Astra's 58.2%. On Artificial Analysis's Intelligence Index it ranks 4th at 55, behind Opus 5.5's 58. Most figures are Moonshot's own and await independent verification.
How much does the Kimi K3.1 Pro API cost?
$2 per million input tokens and $10 per million output tokens ($0.20 cached input) — 33% lower than Kimi K3. The Highspeed tier costs $4/$20 per million for 2–3× faster responses.
Is Kimi K3.1 Pro open source?
Not yet fully: the model runs today only via Moonshot's apps and API. Moonshot has committed to publishing the complete open weights by 4 November 2026, which would make it the world's highest-scoring open-weights model.
Takeaway: Kimi K3.1 Pro compresses the open-vs-closed AI gap to its smallest ever — near-frontier agentic coding at a mid-tier price, with free weights two weeks out. The 4 November open release is the date that actually matters: that's when independent testers find out if Moonshot's numbers hold.
Follow World Tech on TechByCountry for the next AI model update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- Moonshot AI — Introducing Kimi K3.1 Pro: Advancing Open Intelligence (official tech blog, 23 Oct 2026)
- Artificial Analysis — Kimi K3.1 Pro launches at 4th on the Intelligence Index (independent evaluation, 23 Oct 2026)
- madrobot.blog — Kimi K3.1 Pro nearly matches Claude Opus 5.5, weights by Nov 4 (23 Oct 2026)
- GenAInews — Moonshot releases Kimi K3.1 Pro on its API with open weights promised by November 4 (23 Oct 2026)
- OfficeChai — Moonshot AI Releases Kimi K3.1 Pro (23 Oct 2026)
- izz.co.in — Kimi K3.1 Pro Is Here: Moonshot's 2.8T Model Nearly Matches Claude Opus 5.5 (23 Oct 2026)
- Reuters — "Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks" (24 Oct 2026)
Comments
Post a Comment