Kimi K3.1 Pro: near-frontier results, a mid-range price and open weights on the way

Picture a small software agency that builds and maintains web apps for local clients. Its AI bill has crept up as more of the coding, testing and documentation runs through Claude or GPT. The question it keeps asking is whether a cheaper model is good enough for the bulk of the work, with the expensive one saved for the hardest jobs. Kimi K3.1 Pro is the strongest case yet that the answer might be yes.

Moonshot AI introduced Kimi K3.1 Pro on October 23, 2026, the first model in its K3.1 series. Moonshot calls it its most capable model to date in its largest size class, with a 1-million-token context window and substantial gains in coding, 3D creation, financial research and agent work. It launched alongside Kimi K3.1 Pro Highspeed, which Moonshot says runs two to three times faster. Both are available now in Kimi's apps and API, and Moonshot says the full model weights will be released by November 4. ¹

What Moonshot's own tests show

Moonshot published results against the leading closed models. On Terminal-Bench 4.0, a test of coding agents on realistic command-line tasks, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. On BenchCAD, which tests rebuilding 3D objects as CAD code from pictures, it scores 92.7% with tools, close to Opus 5.5 at 96.2% and Astra at 95.9%. ¹

Moonshot also reports a state-of-the-art 64.4 on OfficeQA Pro v2, a test of analyzing real financial documents, and large gains over its previous model on its own finance tests. On its internal factual-accuracy stress test, the rate of false claims fell from 23.4% for Kimi K3 to 8.3%. Those are Moonshot's own benchmarks, run its way, so treat them as direction rather than proof. ¹

What independent testing says

Artificial Analysis runs every model through the same ten evaluations. It scores Kimi K3.1 Pro at 55 on its Intelligence Index, behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56, and ahead of GPT-6 Astra, Gemini 4 Argon and GPT-6.1 Sol, all at 52 to 53. On Artificial Analysis's own run of Terminal-Bench 4.0, Kimi posts 67.2%, the highest of the models in our comparison. ²

Cost is the headline. Artificial Analysis puts Kimi at about $2.12 per index task, roughly a third of Claude Opus 5.5's $5.98 and well under Sonnet 5.5's $5.46, for a score only a few points lower. GPT-6.1 Sol remains the cheapest model in this group at $0.72, but it also scores three points lower. ²

Two numbers temper that. First, speed: Artificial Analysis measured Kimi at about 51 output tokens per second, about half Opus 5.5's 98, so long jobs take noticeably longer. Second, made-up answers: on its AA-Omniscience test, Kimi's hallucination rate is 44%, better than its predecessor's 53% and Opus 5.5's 59%, but well behind Gemini 4 Argon's 15%. ²

Price and where to get it

On Kimi's own API, K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20; the Highspeed version costs twice that. That is the same token price as GPT-6.1 Sol and Gemini 4 Argon's launch price, and half the price of Claude Opus 5.5. ²³

Kimi K3.1 Pro is available in the Kimi app and on kimi.ai, in the Kimi Work desktop app, in Kimi Code and through the Kimi API, where Moonshot offers zero data retention for API customers. Kimi Enterprise admins have to switch it on, since it is off by default. It is also listed on OpenRouter, but for now every request there is served by Moonshot itself. ¹³

Why the November 4 weights release matters

Today, using Kimi K3.1 Pro means sending your data to Moonshot's servers, whichever app or API route you take. That changes when the full weights are released, which Moonshot has promised by November 4. With open weights, other cloud providers can host the model, and larger companies can run it on infrastructure they control, so your prompts and files need never reach Moonshot. ¹³

Until then, treat it like any cloud AI service whose data terms you have not reviewed: keep customer personal data, contracts and anything regulated out of it, use the zero data retention option if you build on the API, and read Moonshot's terms on where data is processed. Its predecessor, Kimi K3, was released under Moonshot's own Kimi K3 license rather than a standard open-source license, so check the license terms that arrive with K3.1 Pro's weights before building a product on it. ¹

A cybersecurity note

Moonshot published an unusually detailed account of how it limited the model's hacking ability before releasing open weights. It says it found and removed cyber-attack-related material from two stages of training. On its internal test of finding and exploiting software vulnerabilities, a version trained without those measures solved 90.5% of problems, while the released model solves 81.0%. Moonshot says it deliberately keeps open-weight models below the frontier of cyber capability. That is a meaningful reduction, not an elimination, and a reminder that open models carry dual-use risk. ¹

Should your business use it?

The broader point is that the gap between the best closed models and the best openly released ones is now only a few points on independent tests. For a business, that means more choice on price and where data lives, as long as someone checks the results on your actual work before switching.