Kimi K3.1 Pro: Moonshot's benchmarks, $2 price and a hallucination claim that needs a caveat

Kimi K3.1 Pro, released by Moonshot AI on October 23, 2026, is a near-frontier model priced at $2.00 per million input tokens and $10.00 per million output tokens, and Artificial Analysis ranks it fourth of 215 models on its Intelligence Index with a score of 55. Moonshot also claims a sharp fall in hallucination, from 23.4 percent to 8.3 percent on its own hard test. Both numbers are real; only one comes from an outsider.

That distinction runs through the whole launch. Here is what is claimed, what is checked and what to wait for.

What Moonshot released

K3.1 Pro is the first model in a new K3.1 series, following K3, which Moonshot released in July 2026 as a 2.8 trillion parameter mixture-of-experts model. It has a 1 million token context window and is available through Kimi.ai, the mobile apps, Kimi Work, Kimi Code and the API as kimi-k3.1-pro, plus a kimi-k3.1-pro-highspeed variant that runs two to three times faster. For enterprise accounts the model is off by default until an administrator enables it. Zero data retention is available for API customers.

Moonshot says full open weights will arrive "by Nov 4." It has not said under what license. Kimi K3 shipped under a modified MIT license with revenue and user thresholds, according to the technology site Noze, so the K3.1 terms are worth reading before anyone builds a product on them. Artificial Analysis currently lists the model as proprietary with weights not publicly available, which fits a model whose weights have been promised but not posted.

What the independent index says

Artificial Analysis reports a score of 55 and a fourth-place ranking among 215 models. It lists the price at $2.00 and $10.00 per million input and output tokens, a blended cost of about $1.54 per million tokens and about $2.11 per Intelligence Index task. Its output speed is 73.5 tokens per second against a median of 82.3. So K3.1 Pro is smart for its price but not fast.

The comparison to K3 is useful. When K3 launched in July, Artificial Analysis placed it fourth of 189 models with a score of 57, at $3 and $15 per million tokens, according to Datanorth. K3.1 Pro scores slightly lower on the index at a third less per token. That is a trade, not a leap.

The vendor benchmarks, read carefully

Moonshot's page leads with coding. On Terminal-Bench 4.0 it reports 65.7 percent for K3.1 Pro, against 66.4 percent for Claude Opus 5.5 and 58.2 percent for GPT-6 Astra. On BenchCAD with tools it reports 92.7, against 96.2 for Opus 5.5 and 95.9 for GPT-6 Astra. In finance it claims a state-of-the-art 64.4 on OfficeQA Pro v2 and gains of 7 to 12 points on its own internal financial modeling and research tests.

Three caveats from Moonshot's own page change how to read these. Many comparisons use internal benchmarks. Competitor scores are sometimes taken from other vendors' reports or leaderboards, with different harnesses. And some external models were found retrieving reference answers online, with those tasks scored zero. On the two coding and design tests where it lists Opus 5.5 head to head, Opus 5.5 is ahead, a fact the launch headline does not stress. For how to read such tables generally, see our guide to what AI leaderboards rank and what a top score hides.

The hallucination claim

Moonshot reports that on its Kimi Hallu Bench Hard, the claim-level hallucination rate with tools falls from 23.4 percent for K3 to 8.3 percent for K3.1 Pro, and that responses with at least one factual error fall to 42.4 percent of K3's rate. In the same table GPT-6 Astra scores 5.5 percent and Claude Opus 5.5 scores 15.0 percent with tools.

Moonshot is explicit that this is a stress test built to expose errors and that the rates "do not represent overall error rates in daily use." That is the right reading. The test is Moonshot's own, so the safest summary is that the lab saw a large improvement on a hard internal measure, with the best closed model still ahead on it.

Cyber capability and the open-weights question

Moonshot says it removed cyber-related data during training and that K3.1 Pro's cyber ability "remains behind" publicly available frontier proprietary models that have been out for at least three months. On its xdev-bench-hard test it reports 81.0 for K3.1 Pro against 100 for GPT-5.6 Sol and 90.5 for an unrestricted Kimi model. That matters because an open-weight model cannot be recalled once released, a point we made in our report on open-weight cyber capability.

Who should care

For most buyers the useful facts are practical. Enterprise administrators must switch K3.1 Pro on, since it is off by default, and API customers can get zero data retention. Artificial Analysis ranks it 113th of 186 on speed and 85th of 113 on cost, so it is a strong performer that is neither the fastest nor the cheapest in its class. If you run agent workloads where reasoning quality matters more than latency, it is worth a test against your own tasks. If you need guaranteed self-hosting today, it is not yet an option.

Our take

K3.1 Pro looks like a real, cheaper near-frontier model, and the independent ranking supports that. It does not look like a leap past the US labs: the two closest head-to-head scores go to Opus 5.5, and the hallucination gain is measured on a test only Moonshot runs. If you use open-weight models, the thing to watch is November 4: whether the weights actually post, and what the license says.