Kimi K3.1 Pro: A Near-Frontier Model You Can Download, Almost

An open model within a point of the frontier

Moonshot AI released Kimi K3.1 Pro on 23 October 2026, together with a faster K3.1 Pro Highspeed variant (Kimi). It is a post-trained version of Kimi K3 with substantially more post-training compute. Moonshot calls it its most capable 3T-class open-weight model, with a 1 million token context window and text, image and video input.

Two things make it worth a production team’s attention. First, on agentic coding it is within a point of Claude Opus 5.5 at half the list price. Second, Moonshot says the full weights will be published by 4 November. If they are, K3.1 Pro becomes the highest-scoring open-weight model on the Artificial Analysis Intelligence Index, three points behind the best proprietary model rather than twelve.

The second point has a large caveat. Moonshot has not published the licence, and until it does, nobody can say whether the weights are usable in a commercial product.

Where it leads, and where it does not

Every row except the Intelligence Index comes from Moonshot’s published benchmark chart, which ran all models at maximum thinking effort (Kimi). The Intelligence Index is independent.

Against K3, Moonshot’s comparisons are on internal benchmarks: Kimi Code Bench 3.0 Long-horizon rises from 55.95% to 61.2%, internal financial research from 58.8 to 70.4, and internal business operations from 8.9 to 37. On the independent Intelligence Index, the score rose from 44 for K3 to 55 (Artificial Analysis).

The pattern is a model that is strong at long-running agent work and office-document tasks, close behind on computer use, and well behind on scientific terminal work and offensive security. The security gap is deliberate, which we come back to below.

Moonshot also reports claim-level hallucinations on its internal Kimi Hallu Bench Hard falling from 23.4% for K3 to 8.3%, against 15.0% for Opus 5.5 and 5.5% for GPT-6 Astra. Moonshot itself calls this “a stress test, not a measure of general error rates”, so treat it as directional. If factual accuracy matters to your product, the evaluation set you already own is a better guide than any vendor benchmark.

Fourth on the Intelligence Index, at a third of Opus’s cost

Vendor tables pick their own comparisons. The Artificial Analysis leaderboard runs the same ten evaluations on every model, including agentic, knowledge-work, coding, science and long-context tasks, so it is the fairest single view of where K3.1 Pro sits.

Source: Artificial Analysis Intelligence Index v4.3.2, default 27-model comparison. Screenshot taken 25 October 2026. Scores are updated as models are re-run, so current values may differ.

K3.1 Pro scores 55, fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Muse Wonder Full has no public API, so among models you can actually call, K3.1 Pro is third. It is ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, all at 53, and of GPT-6.1 Sol at 52.

The open-weight comparison is the bigger story. Before this release, the strongest open-weight model was Xiaomi’s MiMo-V2.6-Pro at 46. If the K3.1 Pro weights ship as promised, the best open-weight model moves from 12 points behind the leader to 3.

Source: Artificial Analysis, Cost per Intelligence Index Task, weighted average cost in USD per task, segmented by token type. Screenshot taken 25 October 2026.

Cost per task is more useful than price per token because it includes how many tokens a model spends reasoning. K3.1 Pro costs $2.11 per Intelligence Index task. Against the models it competes with:

Data from Artificial Analysis as of 25 October 2026. Points per dollar is our calculation, not an Artificial Analysis metric.

Three readings matter for model selection:

• Against Claude, K3.1 Pro is the value option. It gives up 3 points to Opus 5.5 and 1 point to Sonnet 5.5 at roughly 35% and 39% of their cost per task.
• It is not the cheapest route to a score in the low 50s. GPT-6.1 Sol scores 52 for $0.72, a third of K3.1 Pro’s cost, and Gemini 4 Argon matches its value at 53 for $1.99. If your workload does not need the last few points, those are the models to beat.
• Open-weight quality costs more than it used to. MiMo-V2.6-Pro and DeepSeek V4.1 Flash cost $0.13 and $0.26 per task. K3.1 Pro buys about 9 more points for 8 to 16 times the cost. Whether that is worth it depends on the workload, not on the leaderboard.

The price undercuts Opus 5.5 by half

K3.1 Pro prices are from Moonshot’s pricing page. Cache writes are billed separately, at $2.00 per million tokens for a 5-minute TTL and $4.00 for 1 hour on K3.1 Pro. The Opus 5.5 figure is Anthropic’s list price from our September roundup.

Per token, K3.1 Pro is about a third cheaper than K3 and matches GPT-6 Sol. Per task, it is not cheaper at all: Artificial Analysis measures $2.11 per Intelligence Index task for K3.1 Pro against $2.00 for K3. Measure cost per successful outcome on your own workflows, not cost per token. The LLM inference cost guide covers how.

Highspeed is the same model at double the price for two to three times the output speed. It only speeds up generation, not tool calls, so for agent loops dominated by tool latency it buys less than the headline suggests. It suits interactive coding and user-facing chat, where time to the last token is what people notice.

API details that will catch you out

The API is reachable with the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and Zero Data Retention is available for API customers. Several behaviours in Moonshot’s Kimi Code model documentation differ from what most client code assumes:

• Thinking cannot be turned off. Requests with thinking disabled fail with HTTP 400. Any routing layer that disables reasoning for cheap calls needs a per-model exception.
• Three effort levels. reasoning_effort accepts low, high and max. Moonshot’s published benchmarks use max, so expect lower scores and fewer output tokens at the default.
• Typos fail silently in Kimi Code. A mistyped model ID falls back to K2.8 Preview without thinking rather than returning an error. Log the model name the response reports and alert on a mismatch.
• Changing model or effort breaks the cache. Switching model or effort mid-session invalidates the context cache, and Moonshot recommends a new session rather than switching from K3. With cached input at a tenth of the price, a router that flips models per turn will pay for it.
• Use native endpoints. Moonshot recommends its native Responses, Chat Completions or Messages APIs, and warns that proxy protocol translation can cause inconsistent output. If you reach Kimi through a gateway, compare results against the direct API before trusting an evaluation.

Cyber capability was removed on purpose

Moonshot removed cyber-related data from both mid-training and reinforcement learning for the open-weight release, while keeping general-purpose coding data. On its internal xdev-bench-hard exploit benchmark, the checkpoint without these measures solved 90.5% of tasks, and the released model solves 81.0% (The Next Web). Moonshot says its cyber capability remains behind frontier proprietary models that have been public for at least three months.

The reasoning is that open weights cannot be recalled once downloaded. For defenders, this has two consequences. K3.1 Pro is a weaker choice for vulnerability research than the proprietary models, and 81% on an exploit benchmark is still a lot of capability in a model anyone will be able to run. Keep the threat model for your AI applications current, and keep patch windows short for internet-facing services.

Open weights change the decision, if the licence allows

For teams with data-residency or sovereignty constraints, a self-hosted model at this level would change what can run on-premises. Before planning around that, be clear about what is not known yet:

• The licence. Do not assume K3’s terms carry over. Commercial use, attribution requirements and usage thresholds all depend on a document that has not been published.
• The date. “By 4 November” is a commitment, not a release. Do not schedule work that depends on it.
• The hardware. A 3T-class model needs multi-node GPU capacity even when quantised. For most teams, the API will be cheaper than self-hosting below very high volumes. Our GPU inference cost guide shows how to run that comparison.
• Provider risk. Moonshot is a Chinese company. In July the White House accused it of using restricted Nvidia chips and copying an Anthropic model to build K3, without publishing evidence. Whatever the merits, procurement and legal teams in some sectors will ask, and self-hosted weights answer different questions than API access does.

What to do this week

• Add K3.1 Pro to your evaluation runs. Use the API now so you have quality, latency and cost-per-outcome numbers before the weights arrive. Test at both high and max effort.
• Treat model IDs and effort as configuration. Validate the model name the response reports, and keep model and effort fixed per session so the cache keeps working.
• Re-forecast per task, not per token. Price your workloads at $2/$10 using the token counts you measure in your own runs, and compare that with your current model on cost per successful task.
• Decide what the licence must allow. Agree the commercial, attribution and redistribution terms you need with legal now, so you can say yes or no the day it is published.
• Plan self-hosting only after the numbers. If the API evaluation and the licence both pass, size the hardware against your actual traffic before committing to it.

K3.1 Pro may let you run near-frontier quality on infrastructure you control. Whether it does depends on your evaluation results and a licence that has not been written yet. If you want help building the evaluation and routing layer that answers that quickly, talk to Tensorplay about a model readiness review.