Kimi K3.1 Pro

Thirty-second edition of Model Drop. Today: Kimi K3.1 Pro, which Moonshot AI shipped on October 23, 2026, three months after Kimi K3. Same 2.8-trillion-parameter model underneath, a lot more training on top, and it now sits fourth on Artificial Analysis's index for a third less than K3 cost.

Kimi K3.1 Pro is the first model in Moonshot's K3.1 series and, in Moonshot's words, its "most capable 3T-class open-weight model to date." It's a post-trained version of Kimi K3 (2.8 trillion parameters, 104 billion active per token, text and image input) with a 1M-token context window, and it went live the same day on Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the Kimi API. The API price is $2 per million input tokens and $10 per million output, down from K3's $3 / $15, with cached input at $0.20. A second model, Kimi K3.1 Pro Highspeed (kimi-k3.1-pro-highspeed), runs two to three times faster for $4 / $20. Moonshot says the full weights will be released by November 4.

Moonshot's pitch is that more reinforcement learning, run across "hundreds of downstream domains," gave K3.1 Pro gains everywhere, and its table puts the model level with or ahead of the closed frontier on a lot of rows (see ## The benchmarks). The demos lean on long projects: games in Unity, Unreal and Godot, 3D models built by scripting Blender and CAD tools, and an internal test where it fixed 36 errors in a human-walking simulator until the simulated body could walk for 11.2 seconds. The catches: almost every coding score was run inside Anthropic's Claude Code harness, it trails Opus 5.5 and GPT-6 Astra badly on science terminal tasks, Artificial Analysis found it very wordy, and the weights are a promise for now. The launch post also spends a whole section on cybersecurity, where Moonshot says it deliberately removed hacking-related data from training before the open-weight release.

What's new

K3.1 Pro keeps K3's architecture and spends its upgrade on training. Four things set it apart from K3 and from the closed models it's chasing.

• The upgrade is all post-training. Moonshot calls K3.1 Pro "a post-trained version of Kimi K3, with substantially larger post-training compute," and says it rebuilt the RL framework, sandboxes and rollout scheduling so more compute kept paying off. Artificial Analysis puts the result at 55 on its index, up 11 points from K3, with no new base model.

• Hacking skills, removed on purpose. Moonshot found cyber content in its mid-training data and cyber tasks in its RL mix, and stripped both for the open-weight release. On its own exploit test (xdev-bench-hard), the unfiltered checkpoint solved 90.5% of tasks and the shipped model solves 81.0%. Moonshot's stated rule: no open weights that "extend the absolute frontier of cyber capability."

• Fewer made-up facts. On Moonshot's internal hallucination stress test, the share of false claims fell from 23.4% for K3 to 8.3%. That's Moonshot's own test on prompts picked to trip models up, so read it as direction, not a daily error rate.

• A faster sibling, and a lower price. K3's biggest complaint was speed. Highspeed runs two to three times faster at double the price, and the standard model costs a third less than K3 did.

The benchmarks

Moonshot ran Kimi K3.1 Pro at max effort (temperature 1.0, top-p 0.95), mostly inside the Claude Code harness, against Claude Opus 5.5, Fable 5.1 and GPT-6 Astra at max (with fallback), and DeepSeek V4.1 Flash 1017 and GLM-5.3 at max. Rival scores are a mix of Moonshot's own runs and official leaderboards. Moonshot's cost chart also puts K3.1 Pro at 52.8% on Terminal-Bench 4.0 at high effort. Internal Kimi benchmarks and blank cells in Moonshot's table are left out.

Moonshot's numbers

• Terminal-Bench 4.0 (multi-hour terminal work): 65.7%. −0.7 under Claude Opus 5.5 (66.4%), +7.5 over GPT-6 Astra (58.2%), +7.8 over Fable 5.1 (57.9%), +23.4 over DeepSeek V4.1 Flash 1017 (42.3%), +23.9 over GLM-5.3 (41.8%).

• Terminal-Bench Science (science tasks in a terminal): 38.1%. −30.0 under GPT-6 Astra (68.1%), −25.2 under Claude Opus 5.5 (63.3%), −1.9 under Fable 5.1 (40.0%), +22.4 over DeepSeek V4.1 Flash (15.7%), +30.0 over GLM-5.3 (8.1%).

• SWE-Marathon (very long coding jobs): 51.7%. +5.4 over Claude Opus 5.5 (46.3%), +6.1 over Fable 5.1 (45.6%), +9.2 over GPT-6 Astra (42.5%), +9.2 over GLM-5.3 (42.5%), +11.7 over DeepSeek V4.1 Flash 1017 (40.0%).

• DeepSWE v1.1 (fixing real repositories): 75.5%. −1.5 under DeepSeek V4.1 Flash 1017 (77.0%), +1.3 over Claude Opus 5.5 (74.2%), +1.4 over GPT-6 Astra (74.1%), +8.1 over Fable 5.1 (67.4%), +8.6 over GLM-5.3 (66.9%).

• ProgramBench (writing whole programs): 31.5%. −33.5 under Claude Opus 5.5 (65.0%), −22.0 under Fable 5.1 (53.5%), −18.5 under GPT-6 Astra (50.0%), +2.4 over DeepSeek V4.1 Flash 1017 (29.1%), +12.5 over GLM-5.3 (19.0%).

• Humanity's Last Exam (with tools) (expert questions): 70.0%. +2.3 over Claude Opus 5.5 (67.7%), +4.4 over Fable 5.1 (65.6%), +7.5 over GLM-5.3 (62.5%), +11.2 over DeepSeek V4.1 Flash 1017 (58.8%), +12.8 over GPT-6 Astra (57.2%).

• Humanity's Last Exam (no tools) (expert questions, full set): 51.1%. −8.7 under Claude Opus 5.5 (59.8%), −6.5 under Fable 5.1 (57.6%), −2.5 under GPT-6 Astra (53.6%), +8.5 over DeepSeek V4.1 Flash 1017 (42.6%).

• GPQA Diamond (graduate science): 94.2%. −1.9 under GPT-6 Astra (96.1%), +0.3 over Claude Opus 5.5 (93.9%), +0.5 over Fable 5.1 (93.7%), +2.0 over DeepSeek V4.1 Flash 1017 (92.2%), +2.5 over GLM-5.3 (91.7%).

• Agents' Last Exam (real professional tasks): 43.2%. +0.6 over Claude Opus 5.5 (42.6%), +5.5 over DeepSeek V4.1 Flash 1017 (37.7%), +10.4 over GPT-6 Astra (32.8%), +17.0 over GLM-5.3 (26.2%).

• Agents' Last Exam (CLI) (professional tasks in a terminal): 44.2%. +3.5 over Claude Opus 5.5 (40.7%), +5.4 over GPT-6 Astra (38.8%), +8.2 over DeepSeek V4.1 Flash 1017 (36.0%), +14.8 over GLM-5.3 (29.4%).

• OfficeQA Pro v2 (financial documents): 64.4%. +2.2 over Claude Opus 5.5 (62.2%), +9.9 over Fable 5.1 (54.5%), +10.0 over GPT-6 Astra (54.4%), +24.0 over DeepSeek V4.1 Flash 1017 (40.4%), +33.3 over GLM-5.3 (31.1%).

• Frontier Finance (investment workflows): 58.6%. −2.2 under Fable 5.1 (60.8%), +1.1 over GLM-5.3 (57.5%), +1.9 over Claude Opus 5.5 (56.7%), +4.1 over DeepSeek V4.1 Flash 1017 (54.5%), +13.6 over GPT-6 Astra (45.0%).

• GDPval-AA (paid office tasks): 1846 Elo. −20 under Claude Opus 5.5 (1866), +92 over Fable 5.1 (1754), +100 over DeepSeek V4.1 Flash 1017 (1746), +196 over GLM-5.3 (1650), +273 over GPT-6 Astra (1573).

• AA-Briefcase v1.1 (office deliverables): 1678 Elo. −125 under Claude Opus 5.5 (1803), +6 over Fable 5.1 (1672), +108 over DeepSeek V4.1 Flash 1017 (1570), +109 over GPT-6 Astra (1569), +171 over GLM-5.3 (1507).

• AutomationBench-AA (business automations): 73.3%. +3.8 over Claude Opus 5.5 (69.5%), +4.8 over GPT-6 Astra (68.5%), +6.7 over DeepSeek V4.1 Flash 1017 (66.6%), +11.1 over GLM-5.3 (62.2%), +13.9 over Fable 5.1 (59.4%).

• APEX-Agents (consulting and banking tasks): 67.2%. −6.3 under Claude Opus 5.5 (73.5%), −1.4 under Fable 5.1 (68.6%), +2.5 over GPT-6 Astra (64.7%), +3.8 over DeepSeek V4.1 Flash 1017 (63.4%), +10.6 over GLM-5.3 (56.6%).

• OSWorld-V2.1 (using a computer): 79.6%. −2.2 under Claude Opus 5.5 (81.8%), −1.1 under Fable 5.1 (80.7%), +3.8 over GPT-6 Astra (75.8%), +10.3 over DeepSeek V4.1 Flash 1017 (69.3%).

• BrowseComp (finding facts on the web): 92.1%. +0.6 over GPT-6 Astra (91.5%), +1.3 over Claude Opus 5.5 (90.8%).

• GDP.pdf (reading business PDFs): 34.6%. +0.4 over GPT-6 Astra (34.2%), +4.0 over Claude Opus 5.5 (30.6%), +7.0 over Fable 5.1 (27.6%), +17.4 over DeepSeek V4.1 Flash 1017 (17.2%), +23.4 over GLM-5.3 (11.2%).

• BenchCAD (with tools) (3D models from images): 92.7%. −3.5 under Claude Opus 5.5 (96.2%), −3.2 under GPT-6 Astra (95.9%), +0.1 over Fable 5.1 (92.6%), +11.1 over DeepSeek V4.1 Flash 1017 (81.6%).

• MMMU-Pro (images plus reasoning): 84.1%. −3.6 under Claude Opus 5.5 (87.7%), −2.8 under GPT-6 Astra (86.9%), −2.2 under Fable 5.1 (86.3%), +3.5 over DeepSeek V4.1 Flash 1017 (80.6%).

• ZeroBench (with tools) (hard visual reasoning, pass@5): 63.0%. +2.0 over Claude Opus 5.5 (61.0%), +4.0 over Fable 5.1 (59.0%), +5.0 over GPT-6 Astra (58.0%), +12.0 over DeepSeek V4.1 Flash 1017 (51.0%).

• Chartography (with tools) (reading charts): 90.2%. −0.8 under GPT-6 Astra (91.0%), +1.2 over Claude Opus 5.5 (89.0%), +1.8 over Fable 5.1 (88.4%), +7.2 over DeepSeek V4.1 Flash 1017 (83.0%).

• BabyVision (with tools) (basic visual puzzles): 95.4%. −2.3 under Claude Opus 5.5 (97.7%), −1.2 under GPT-6 Astra (96.6%), +0.3 over Fable 5.1 (95.1%), +2.9 over DeepSeek V4.1 Flash 1017 (92.5%).

• Harvey Legal Agent Benchmark (legal tasks, all criteria passed): 23.7%. +2.7 over DeepSeek V4.1 Flash 1017 (21.0%), +4.6 over Fable 5.1 (19.1%).

• ExploitBench (exploiting real bugs): 74.7%. −25.3 under GPT-6 Astra (100%), −16.3 under Claude Opus 5.5 (91.0%), −8.3 under Fable 5.1 (83.0%), +20.3 over GLM-5.3 (54.4%).

Independent numbers

Artificial Analysis tested Kimi K3.1 Pro at its default max effort, with Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 at max (with fallback), and GPT-6 Astra and Kimi K3 at max.

• Artificial Analysis Intelligence Index (current v4.3.2): 55. +11 over Kimi K3 (44), −3 under Claude Opus 5.5 (58), −1 under Claude Sonnet 5.5 (56), +2 over Fable 5.1 (53), +2 over GPT-6 Astra (53), +2 over Gemini 4 Argon at high (53).

See how it ranks. Every score above is on the Handy AI benchmarks board, where you can see where Kimi K3.1 Pro lands against every model we track, open and closed.

First impressions

The model is a few hours old. Reactions come from X and the Hacker News launch thread (281 points and 150 comments at the time of writing).

The positives

Artificial Analysis (@ArtificialAnlys), which runs its own tests on every major model, had it fourth overall within half an hour of launch.

Fourth on an independent index, ahead of GPT-6 Astra and Fable 5.1, is the number that makes Moonshot's own table easier to believe.

Nathan Lambert (@natolambert), a researcher who writes about post-training, read the benchmarks as a training story.

That matches what Moonshot says it did: same base model as K3, much more reinforcement learning.

pcwelder on Hacker News had K3's biggest problem in mind.

Biggest gripe with K3 was slow responses, partly due to inefficient thinking. So far K3.1 pro is almost twice as fast as k3 on our benchmark. (comment)

One team's test, but speed was the main reason people put K3 down, and Artificial Analysis also clocks K3.1 Pro faster than GPT-6 Astra and Fable 5.1.

The negatives

Lisan al Gaib (@scaling01), who builds his own benchmarks, flagged the test setup.

Moonshot's footnotes confirm it: Terminal-Bench, SWE-Marathon, DeepSWE, ProgramBench and more ran in Claude Code, which isn't how most people will run Kimi. The harness can move scores a lot, so watch for runs in the standard harnesses.

johnmlussier on Hacker News hit the limits on Moonshot's own coding plan.

If you are expecting any sort of usable limits on their subscription plan I suggest adjusting your expectations. I subscribed to $99 a month plan and within 47 minutes with one session (not asking for swarms or anything) was limited by 5-hour quota. (comment)

Artificial Analysis measured 150 million output tokens to run its index, against a median of 87 million, so a model this wordy eats a quota fast.

Hany (@hany99dev), a general X user, didn't buy the "caught up" framing.

Moonshot's own table backs him up on computer use, and Opus 5.5 keeps a small lead on Terminal-Bench too, so "near Opus" is fair and "past Opus" isn't.

Jake's take

This is the closest an open model has come to Opus 5.5, and it's priced like Sonnet 5.5. Fourth on Artificial Analysis, a point behind Sonnet at the exact same $2 / $10, and ahead of Fable 5.1 and GPT-6 Astra. In July I said Moonshot had finally built a model that competes on capability instead of price. K3.1 Pro does both, and it does it by training the same base model harder, which is a little humbling for everyone spending billions on new ones.

Two things bug me. Moonshot ran nearly every coding score inside Claude Code, so the table is partly a measure of Anthropic's harness. And the API docs went live with "[xx.x%]" placeholders where the Terminal-Bench numbers should be, plus hallucination numbers that don't match the blog. Small stuff, but it's the kind of small stuff that makes me wait for the independent runs before getting excited about a 65.7.

The cyber section is the most interesting thing in the launch. A lab saying "here's the checkpoint we didn't ship, here's how much hacking we trained out, and here's why" is more paperwork than some US labs give us. Now ship the weights by November 4 and let people check.