Kimi K3.1 Pro: Moonshot's Near-Frontier AI Model

Updated: October 2026 · 6 min read · By the CreatorSprout team

Launch coverage based on Moonshot AI's announcement and independent benchmark reporting — we have not tested this model hands-on. Vendor figures and independent measurements are kept separate below, because they tell different stories. Pricing verified October 2026; confirm on official pages before you commit. No affiliate links in this article.

That "if" is doing a lot of work, and so is the gap between Moonshot's own benchmark table and what independent evaluators measured. Here's both sides, kept apart.

What Kimi K3.1 Pro is

Kimi K3.1 Pro is a mixture-of-experts language model with 2.8 trillion total parameters, 104 billion active per token, accepting text and image input with a 1-million-token context window. It is the first model in Moonshot's K3.1 series: not a new base model, but a heavy post-training round on Kimi K3, with Moonshot saying it ran far more compute in that phase than on past projects and upgraded its reinforcement-learning system, long-context inference, and code-execution sandboxes.

A second variant launched alongside it: Kimi K3.1 Pro Highspeed (kimi-k3.1-pro-highspeed), which Moonshot says runs two to three times faster. Both are live now in the Kimi app, Kimi Work, Kimi Code, and the Kimi API under the identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed.

Moonshot's demos lean on long, multi-step work: games built in Unity, Unreal, and Godot; 3D models produced by scripting Blender and CAD tools; and an internal test where the model fixed 36 errors in a human-walking simulator until the simulated body could walk for 11.2 seconds. The through-line is agentic coding and office work, not chat.

What Moonshot claims

On Moonshot's own benchmark table, K3.1 Pro sits level with or just behind the closed frontier:

Terminal-Bench 4.0 (agentic coding): 65.7% — versus 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra
BenchCAD (3D model coding with tools): 92.7% — versus 96.2% for Claude Opus 5.5
OfficeQA Pro v2 (office work): 64.4 — versus 62.2 for Claude Opus 5.5, a rare row where Moonshot claims the lead
Internal hallucination stress test: 8.3% claim-level error rate — versus 23.4% for Kimi K3, roughly a two-thirds drop

Two warnings apply to this whole table. First, it is Moonshot's own harness and partly internal tests — the company itself notes its stress-test rates "do not represent overall error rates in daily use." Second, reporting notes that almost every coding score was run inside Anthropic's Claude Code harness, which is an unusual setup choice for a vendor table. Treat these as the company's best case, not as settled fact.

What independent testing found

Independent evaluator Artificial Analysis scored K3.1 Pro at 55 on its Intelligence Index (55.1 in the detailed table), ranking it fourth of 215 models — behind Claude Opus 5.5 (57.6), Claude Sonnet 5.5 (56.0), and Muse Wonder Full (55.8). That is three points off the leader: close, but not tied for the top.

The independent detail is more mixed than Moonshot's table:

Terminal-Bench 4.0: 67% — actually higher than Moonshot's own 65.7%, and the highest Artificial Analysis had measured, versus 60% for Opus 5.5 at its maximum setting. But note the settings differ between the two tables, so the numbers are not directly comparable.
GDPval-AA (simulated knowledge work): 1,846 Elo — behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857.
Humanity's Last Exam: 52% — versus 47% for K3 (a real gain) but 61% for Opus 5.5 (a real gap).
Hallucination (AA-Omniscience): 44% — down from 53% for K3, but far above the 15% reported for Gemini 4 Argon on the same evaluation. The "65% hallucination drop" only holds on Moonshot's own stress test.

Artificial Analysis also flagged that the model is very wordy, and estimated its cost at $2.11 per Intelligence Index task versus $5.98 for Opus 5.5 at maximum effort — a benchmark-specific estimate, not a promise about your workload.

Pricing: cheaper than K3, cheaper than the frontier

API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 — a third below Kimi K3's $3/$15. The Highspeed variant costs $4/$20 for two to three times the speed.

The honest way to read that price: it is genuinely cheap for a model scoring 55 on the Intelligence Index, and the per-task cost estimate backs that up. But price per token is not price per outcome — a wordy model that takes more tokens to finish a task can erase a per-token discount. If you adopt it, measure cost per completed task on your own workload, not per million tokens.

The open-weights promise: due November 4

Moonshot says the full weights will be released by November 4, 2026. No license has been announced yet. Until the weights actually appear, K3.1 Pro is an API model with an open-weight promise — not an open-weight model you can download today. That distinction matters: teams planning to self-host should wait for the actual release and license terms before committing infrastructure or budget.

One reason to take the promise seriously: K3 followed the same pattern in July 2026, with weights released 11 days after the API launch. One reason to stay cautious: a promised release is still a promise, and the license terms — which decide whether commercial self-hosting is actually allowed — are unknown.

Moonshot also used the launch post to address safety directly, saying it deliberately removed hacking-related training data before the open-weight release so the published weights would not advance offensive hacking capability. That is a vendor claim, not an independently verified fact, but it is a notable thing for a lab to put in writing ahead of an open release.

How to try it

Easiest: the Kimi app. The model is in the Kimi picker on the website and apps, plus Kimi Work and Kimi Code.
Developers: the Kimi API. Use kimi-k3.1-pro for the standard model or kimi-k3.1-pro-highspeed for the faster variant.
Start with long agentic work. That is where both Moonshot's and the independent numbers are strongest — multi-step coding, office documents, and spreadsheet-style tasks.
Measure cost per finished task. Track tokens per completed job, not just the $2/$10 headline rate, before comparing it with your current model.
Wait for November 4 before planning self-hosting. The weights and license are not out yet.

The honest caveats

We haven't tested it. Every performance number in this article comes from Moonshot or Artificial Analysis, not from our own runs.
The vendor and independent numbers disagree on hallucination. Moonshot's internal test shows a ~65% drop; the independent test shows 53% to 44%. Believe the independent one for real-world expectations.
It trails the frontier on hard knowledge tests. 52% on Humanity's Last Exam versus 61% for Opus 5.5 is a meaningful gap, not a rounding error.
The weights are a promise, not a download. No license announced, no repository yet. November 4 is the date to watch.
Wordiness costs money. Artificial Analysis flagged verbose outputs; at $10 per million output tokens, long answers add up.

Verdict

Kimi K3.1 Pro is the most interesting open-weight-adjacent launch of the week — not because it beats the frontier (it doesn't, on independent tests), but because it gets within three Intelligence Index points of the leader at a third of its predecessor's price, with downloadable weights promised by November 4. The honest move: try it now through the app or API on your own long coding and office tasks, measure cost per finished job, and hold any self-hosting plans until the weights and license actually appear.

FAQ

What is Kimi K3.1 Pro?

A language model launched October 23, 2026 by Moonshot AI — a post-trained upgrade of Kimi K3 with 2.8 trillion total parameters (104 billion active per token), a 1-million-token context window, and text plus image input. It is available through the Kimi app, Kimi Work, Kimi Code, and the Kimi API.

How much does Kimi K3.1 Pro cost?

$2 per million input tokens and $10 per million output tokens on the API (cached input $0.20). The Highspeed variant costs $4/$20 and runs two to three times faster, according to Moonshot.

Are the open weights available?

Not yet. Moonshot says the full weights will be released by November 4, 2026, and no license has been announced. Until then it is usable only through Moonshot's apps and API.

How good is Kimi K3.1 Pro?

Artificial Analysis ranks it fourth on its Intelligence Index at 55, behind Claude Opus 5.5 (57.6), Claude Sonnet 5.5 (56.0), and Muse Wonder Full (55.8). Moonshot's own table puts it closer to the frontier on coding and office benchmarks, but those are vendor-run figures.

Is Kimi K3.1 Pro better than Claude Opus 5.5?

No, on the independent evidence. It trails Opus 5.5 on the Intelligence Index, on Humanity's Last Exam (52% vs 61%), and on simulated knowledge work — while beating it on one office benchmark in Moonshot's own table and costing far less per task in Artificial Analysis's estimate.