Moonshot AI put Kimi K3.1 Pro on its API on October 23, and its launch post calls the model its "most capable 3T-class open-weight model to date," with full weights promised by November 4. It is a post-trained Kimi K3, and by Moonshot's own chart it trails Claude Opus 5.5 by 0.7 points on Terminal-Bench 4.0 while charging half as much per token. The vendor number is where the story starts, not where it ends.
Seven tenths of a point, by Moonshot's arithmetic
The headline figure is 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. The footnote matters more than the bar. Moonshot ran its own model in the Claude Code harness at maximum reasoning effort; the rival scores are the best each vendor published or the official leaderboard shows, not Moonshot re-runs. Same exam, different proctors.
Elsewhere the launch table is mixed, as honest ones are. K3.1 Pro leads on OfficeQA Pro v2 at 64.4 and trails Opus 5.5 on BenchCAD with tools, 92.7 to 96.2. VKTR's tabulation of the same vendor numbers shows wider gaps to Opus 5.5 on Program Bench (31.5 against 65.0) and Terminal-Bench-Science (38.1 against 63.3).
Price is where it stops being close. K3.1 Pro lists at $2 per million input tokens and $10 per million output, a third below K3's $3 and $15; Opus 5.5 costs $4 and $20. Moonshot's docs put the K3 series' context at 1,048,576 tokens, and the K3.1 Pro quickstart allows a completion budget up to that same figure, defaulting to 131,072. A Highspeed variant runs two to three times faster at twice the price.
What the independent graders saw
Two outside harnesses bracket the claim. Vals.ai, calling Moonshot's API, measured 63.64% on Terminal-Bench 4.0, third on its board and 1.5 points behind its own Opus 5.5 run at 65.15%. Artificial Analysis's launch-day run, as OfficeChai reported it, put K3.1 Pro at 67.2%, the top score on that board. On Vals's runs, K3 had managed 17.17%. Post-training did most of the work.
The broader index shifted after launch. At launch, Artificial Analysis's Intelligence Index v4.3 scored K3.1 Pro 55, fourth, behind Opus 5.5 at 58 and ahead of GPT-6 Astra and Gemini 4 Argon at 53. On October 27 the firm moved to v5, adding Cognition's private FrontierCode 2.0 and leaning harder on private test sets. K3.1 Pro now scores 46 to Opus 5.5's 58, and the open-versus-closed gap the firm tracks widened from 3 points to 12. Same model, new ruler. It still leads the best released open models, DeepSeek V4.1 Flash and GLM-5.3 at 32, by 14 points.
Cost per task survives the re-measure better than rank does. A v5 index task costs $3.68 on K3.1 Pro against $10.04 on Opus 5.5 and $6.24 on GPT-6 Astra, even though K3.1 Pro emitted 110 million output tokens on the suite against a 70 million median. It is also slow: 35.1 tokens per second on Moonshot's API, 192nd of 194.
Hallucination shows why definitions matter. On Moonshot's internal adversarial test, the claim-level rate fell from 23.4% for K3 to 8.3%, against 15.0% for Opus 5.5 and 5.5% for GPT-6 Astra, on a test Moonshot says does not reflect daily use. Artificial Analysis's Omniscience benchmark, which defines the rate differently, moved from 53% to 44%. Same direction, different planet.
Open weights, heavy racks
Open here means downloadable, not portable. Moonshot's model list puts K3 at 2.8 trillion parameters, and K3.1 Pro is the same network post-trained; Latent Space reports 104B active per token. K3's weights were 1.56TB on Hugging Face, stored as MXFP4 weights with MXFP8 activations. At 80GB per H100, that fills about 20 cards before a single token of KV cache; a community estimate on the model's Hugging Face page lands at 21 to 22 H100s, or $550,000 to $660,000 at list. Openzeka, an Nvidia distributor, served K3 on one eight-GPU DGX B300 at 83 to 101 tokens per second for a single user.
Read the license before the download. K3's license dropped the "modified MIT" label and requires a separate agreement from any model-as-a-service operator with more than $20 million in revenue over 12 months. Moonshot has not said what K3.1 Pro will ship under. It has said the open release is deliberately blunted on offense: removing cyber data from mid-training and RL cut its internal exploit benchmark from 90.5% to 81.0%, still far above K3's 33.3%.
An allegation, and a border
The model arrives with an allegation attached. On October 12 Anthropic published its distillation defenses, naming Moonshot and Zhipu and alleging that requests to Kimi models "were secretly forwarded to Claude" for distillation. Its September threat report, as Bloomberg covered it, cited nearly 300,000 customer requests sent mainly to Opus through 5,380 fraudulent accounts. These are Anthropic's allegations; no court or regulator has ruled on them.
Moonshot declined to comment on the September report, and China's Commerce Ministry has said US distillation accusations have no factual or legal basis. In July, answering a separate White House accusation, Moonshot's business head denied distilling and credited K3's gains to its own architecture. We found no Moonshot response to the October post.
Beijing adds a border to the story. China's October 16 draft would let regulators restrict exports of models with 2 trillion or more parameters, and AP names Kimi K3 among them; technology already public through open-source release is excluded. Whether publishing K3.1 Pro's weights by November 4 counts as an export or as entering the public domain is a question for the comment period, which closes November 15.
The 30,000-foot read
On the benchmark agentic-coding buyers watch most, a model bound for open release now sits within two points of Opus 5.5 on two counts and on top of a third, at a little over a third of Opus's cost per index task. On broader measures it is a tier down, and the ruler moved four days after launch. Both are true. Only one fits in a headline.
What deserves attention is everything outside the score. A 1.56TB checkpoint is open the way a cathedral is open: anyone may enter, few can pay for the roof. For most teams K3.1 Pro will be a cheap, slow, verbose API with an unannounced license and an unresolved provenance dispute; for clouds and sovereign deployments, perhaps more. Watch November 4 for the weights and license, then for a neutral-harness Terminal-Bench run and Beijing's treatment of open releases above two trillion parameters. Your own eval suite remains the one benchmark nobody can reweight on you.
Sources: Moonshot AI, Kimi K3.1 Pro launch post (Terminal-Bench 4.0, OfficeQA, BenchCAD, hallucination and cyber figures, harness footnote, Nov 4 weights promise); Kimi API model list (K3 at 2.8T parameters); Kimi API pricing page, Chinese edition (1,048,576-token K3-series context); Kimi K3.1 Pro quickstart (completion-token limits); Simon Willison (Oct 23 release, $2/$10 and Highspeed pricing, frontier comparison); Simon Willison on Kimi K3 (1.56TB weights, K3 license terms); Latent Space (104B active, K3's prior pricing); Claude Opus 5.5 docs (Opus 5.5 pricing); VKTR (Program Bench and Terminal-Bench-Science vendor figures); Vals.ai (independent Terminal-Bench 4.0 runs for K3.1 Pro, K3 and Opus 5.5); OfficeChai (Artificial Analysis launch-day v4.3 index, Terminal-Bench and Omniscience figures); Artificial Analysis, Intelligence Index v5 (v5 scores, gap, cost per task); Artificial Analysis model page (cost per task, verbosity, speed); Openzeka (K3 weight format, DGX B300 throughput); Hugging Face discussion (community hardware and cost estimates); Anthropic, distillation defenses (Oct 12 allegation naming Moonshot and Zhipu); Bloomberg Government (September report figures); Bloomberg Línea (Moonshot declined comment, Commerce Ministry response); Implicator (Moonshot's July denial); AP via ITV (2T-parameter criterion, K3 above it); Xinhua (open-source exclusion). Caveats and corrections: all Moonshot comparison figures are vendor-reported; only K3.1 Pro's own Terminal-Bench run was Moonshot's (in the Claude Code harness), while rival scores are the maxima published by each vendor or the official leaderboard, not Moonshot re-runs. Correction to launch-week reporting: the widely cited Artificial Analysis score of 55 was the v4.3 figure; since v5 launched Oct 27 the score is 46. Artificial Analysis's 67.2% Terminal-Bench figure comes from OfficeChai's summary and did not appear on Artificial Analysis's model page when we checked. The 8.3% hallucination rate is from Moonshot's internal adversarial benchmark. Moonshot documents 2.8T parameters for K3 and calls K3.1 Pro a post-trained K3 but does not state K3.1 Pro's count directly; the 104B active figure is secondary, and a July Latent Space headline put K3 at 50B active, so treat the active count as unsettled. Self-hosting figures describe K3, since K3.1 Pro's weights, license and hardware guidance were unpublished at writing; Hugging Face hardware and cost numbers are community estimates, and the 20-card figure is our arithmetic. Moonshot's English pricing page and The Decoder's report could not be opened by our fetcher, and Moonshot's Chinese price tables did not render, so prices rest on Simon Willison and Latent Space, which agree with Vals.ai and Artificial Analysis. Moonshot's July denial is reported by Implicator via Asia Times and answered the White House accusation, not Anthropic's October post. The Commerce Ministry response is as reported by Bloomberg Línea.