Moonshot's Kimi K3.1 Pro Scores 55 on Artificial Analysis at $2/$10, Weights Due Nov 4

TL;DR

Kimi K3.1 Pro from Moonshot AI went live on October 23 across the Kimi apps, Kimi Code and the API. Artificial Analysis scores it 55 on its Intelligence Index, fifth on the leaderboard and behind only Claude Opus 5.5, Sonnet 5.5 and Meta's Muse Wonder Full. Every model above it is closed. API pricing is $2 in and $10 out per million tokens, down from Kimi K3's $3/$15, and Moonshot says the full weights will be released by November 4, 2026. Until they land, this is an API model with a promise attached.

What shipped

K3.1 Pro is a post-trained version of Kimi K3, the model Moonshot released in July. Moonshot's announcement says it got "substantially larger post-training compute" and calls it "the most capable 3T-class open-weight model to date." The context window stays at 1 million tokens. Moonshot did not publish a new architecture breakdown or active-parameter count.

There are two API model IDs: kimi-k3.1-pro and kimi-k3.1-pro-highspeed. According to Moonshot and Simon Willison's write-up, the Highspeed variant runs 2-3x faster at twice the price, though cache-hit input costs the same on both. You get three reasoning levels: low, medium and max. Zero Data Retention is available to API customers, and in Kimi Code you switch to it with /model.

Where it lands on the leaderboard

Artificial Analysis ranks K3.1 Pro fifth on its model leaderboard. The four entries above it are Claude Opus 5.5 at max effort (58), Claude Sonnet 5.5 at max (56), Opus 5.5 at xhigh (56) and Muse Wonder Full (56). GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 all score 53, and GPT-6.1 Sol scores 52.

That makes K3.1 Pro the highest-scoring model on the board with announced open weights. Willison put it this way: "This is the first time a (soon-to-be) open weight model has scored this highly against the frontier." Kimi K3, the model it replaces, scored 44, so one round of post-training bought eleven points.

The price, and the catch in the price

At $2/$10 per million tokens, K3.1 Pro matches GPT-6.1 Sol's list price and charges half of Opus 5.5's $4/$20. Artificial Analysis lists a 90% cache discount and a blended rate of $1.54 per million tokens. It also measured 77.1 output tokens per second, about twice the 38.5 it clocked for Kimi K3.

List price is only half the bill. Artificial Analysis flags the model as "very verbose": it generated 150 million output tokens to finish the Intelligence Index, against a median of 87 million. Per task, K3.1 Pro costs $2.11. That is well under Opus 5.5 max ($5.98) and GPT-6 Astra ($3.26), roughly level with Gemini 4 Argon ($1.99), and nearly three times GPT-6.1 Sol ($0.72), which scores three points lower. The cheap tokens are real. So is the model's habit of using a lot of them, like a contractor with a low hourly rate who bills for every phone call.

Moonshot's own numbers

The benchmarks below come from Moonshot's announcement, run at max reasoning effort, temperature 1.0 and top-p 0.95. Several are internal suites, so treat them as vendor claims until outside testers reproduce them.

• Terminal-Bench 4.0: 65.7%, against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra.
• Kimi Code Bench 3.0 long-horizon (internal): 61.2%, up from 55.95% for K3.
• BenchCAD with tools (voxel IoU): 92.7%, against 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra.
• Kimi Hallu Bench Hard with tools: claim-level hallucination rate down from 23.4% on K3 to 8.3%. On the same test Moonshot reports 5.5% for GPT-6 Astra and 15.0% for Opus 5.5.
• Internal business operations suite: 37, up from 8.9 for K3.

The Terminal-Bench figure is the one coding-agent builders should check first, since it is a public benchmark and the gap to Opus 5.5 is under a point. The hallucination figures are internal, and Moonshot wrote the test.

The cyber trim on the open weights

Moonshot says it removed cyber-related data from both mid-training and RL for the open-weight release. It published the cost. On its internal xdev-bench-hard (pass@3), K3.1 Pro scores 81.0, while a checkpoint trained without that risk reduction scored 90.5. Moonshot says K3.1 Pro's cyber capability stays behind frontier proprietary models that have been public for at least three months. After a month in which AI-found exploits kept making the news, a lab that publishes the score it gave up for safety is doing something unusual.

What to do with it

For API users, the switch is cheap to test. Willison ran it through the OpenAI-compatible endpoint at https://api.moonshot.ai/v1 with his LLM command-line tool and the --responses flag. His usual SVG pelican-on-a-bicycle at max effort came out "pretty solid," he wrote, though both of the pelican's legs ended up on the same side of the frame.

For self-hosters, the date to watch is November 4. Moonshot has not published a license for K3.1 Pro, so read it before building on the weights. A 3T-class checkpoint will not fit on a homelab card either, so most people will reach it through third-party hosts or heavy quantization.

The Kimi app also gained 3D Make, which generates printable 3D models, and Moonshot says the model can drive Unity, Unreal, Godot, Blender and CAD scripting without MCP connectors. Those claims rest on Moonshot's own demos for now.

Key Takeaways

• Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fifth overall, behind Opus 5.5 max (58) and ahead of GPT-6 Astra, Gemini 4 Argon and Fable 5.1 (53).
• No model with announced open weights has scored higher, and Moonshot says the full weights ship by November 4, 2026, with no license published yet.
• API pricing is $2/$10 per million tokens, a third below Kimi K3 and half of Opus 5.5, with a Highspeed variant at 2-3x speed for double the price.
• Heavy verbosity (150M output tokens on the index versus an 87M median) puts real cost at $2.11 per task, near Gemini 4 Argon and far above GPT-6.1 Sol.
• Moonshot reports 65.7% on Terminal-Bench 4.0, within a point of Opus 5.5, but most of its other headline numbers come from internal suites.
• The open-weight release has cyber training data stripped out, and Moonshot published the cost: 81.0 versus 90.5 on its internal exploit benchmark.

Sources: Moonshot AI: Kimi K3.1 Pro, Artificial Analysis: Kimi K3.1 Pro, Artificial Analysis leaderboard, Artificial Analysis: Kimi K3, Simon Willison: Introducing Kimi K3.1 Pro, Kimi API pricing, Kimi K3 on Hugging Face