Cheaper tokens, a coding chart that sits within a point of Claude Opus 5.5, and a system you still cannot run yourself. That is the offer in front of a buying committee. The useful question is which part of it your workflows can actually survive.
Moonshot AI put Kimi K3.1 Pro on an API on October 23, 2026, at $2 per million input tokens and $10 per million output tokens. R&D World reports those rates are down from $3 and $15 on Kimi K3, which launched in July, and half the price of Claude Opus 5.5 at $4 and $20. They match the sticker on GPT-6.1 Sol, which launched on September 29. R&D World credits that comparison to an October 23 note from Simon Willison. The weights were not in the box. Moonshot has said it will publish them by November 4. The license is still blank. Until both exist, Artificial Analysis, as cited by R&D World, lists the model as proprietary.
A heavier training pass, not a new model
K3.1 Pro is the first model in the K3.1 series. It is a further training pass on Kimi K3, not a new design. Moonshot says it spent much more compute on that later stage, including joint reinforcement learning across what it calls hundreds of downstream domains. It has not said what those domains actually cover.
The size in Moonshot’s launch post is the K3 size: 2.8 trillion total parameters, 104 billion active, and a 1-million-token context window. The K3 technical report (arXiv:2607.24653) is where the architecture is described, and nothing in the later reporting shows those figures independently remeasured. The same report says routing turns on 16 of 896 routed experts. That fraction is not the same as 104 billion out of 2.8 trillion (about 1.8 percent of experts versus about 3.7 percent of parameters), so the idle-headcount arithmetic does not hold. For anyone pricing a cluster, the ordinary mixture-of-experts fact still applies. A slice of the model runs on each token, and you still have to store the whole thing.
K3’s weights were released. K3.1 Pro is API-first. Open-weight, here, is a commitment with a date on it. There is no download yet. November 4 is Moonshot’s stated deadline, as reported by R&D World. The release has not happened.
You can call it now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, including an OpenAI-compatible endpoint. The documented model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The quickstart lists text, image, and video inputs. Images are not accepted as public URLs. The documented path is base64 or an uploaded file ID.
The sticker is not the invoice
The half-price story is the standard tier only. Highspeed is $4 in and $20 out, the same sticker R&D World assigns to Opus 5.5. Moonshot says that variant runs two to three times faster, and has not said on which workload. Moonshot’s rate card, as carried in the launch coverage, lists cache hits at $0.20 per million input tokens. R&D World confirms that Highspeed does not change the cached-input rate relative to the standard tier, but it does not restate the $0.20 figure.
The knobs are narrower than most teams expect. Thinking stays on. You can set reasoning effort to low, high, or max, and max is the default. Temperature is fixed at 1.0 and top-p at 0.95. Output defaults to 131, 072 tokens and can be raised to 1, 048, 576. That cap is a different limit from the 1-million-token context window. If your integration depends on turning the temperature down, this API will not indulge you. Effort is the dial.
Treat one habit as operating guidance until you have tested the opposite. The docs say to send the full assistant message back on the next turn, prior reasoning included. On K3, Moonshot warned that dropping that history makes the output unstable. Long-running agents that trim transcripts to save money may be buying a cheaper prompt and a worse answer.
Read the method, then the bars
Moonshot’s own sheet, restated by R&D World, is where the flattering numbers live, and the method belongs in front of them. K3.1 Pro was run at maximum reasoning effort, temperature 1.0 and top-p 0.95, mostly through Claude Code, an agent setup that drives a coding terminal. Rival scores on that sheet mix Moonshot’s runs, vendor system cards and blogs, and public leaderboards. Claude results may include fallbacks. GPT results may include cyber safeguards. Treat the sheet as the vendor’s argument. It is not a confirmed sweep.
On Terminal-Bench 4.0, that sheet puts K3.1 Pro at 65.7, Opus 5.5 at 66.4, and GPT-6 Astra at 58.2. On SWE-Marathon, K3.1 Pro leads the sheet at 51.7, ahead of Opus 5.5 at 46.3. The same comparison also includes Claude Fable 5.1, DeepSeek V4.1 Flash, and Zhipu’s GLM-5.3. Those two coding results are specific enough to test. They do not amount to a general claim of parity.
Then the floor drops. On Terminal-Bench Science 0.1, K3.1 Pro scores 38.1. The same sheet puts GPT-6 Astra at 68.1, Opus 5.5 at 63.3, and Fable 5.1 at 40.0. That is about 25 points behind Opus and 30 behind Astra. If the work looks like scientific computing in a terminal, this chart is a warning label, not a trophy.
The Astra figures are already in conflict. Comparison notes published with the launch cite OpenAI’s own pages at 57.9 on Terminal-Bench 4.0 and 64.6 on the science bench, against 58.2 and 68.1 on Moonshot’s chart. R&D World repeats the chart and does not mention those pages. The disagreement is unresolved.
One DeepSeek figure needs the same care. R&D World says the DeepSeek science score on Moonshot’s sheet came from an older model version. A later Moonshot run of the current version scored 9.3 on Terminal-Bench Science 0.1. An older listing and 9.3 are not the same result.
On an internal hallucination stress test, Moonshot reports claim-level errors falling from 23.4% on K3 to 8.3%. That is still worse than the 5.5% it lists for GPT-6 Astra on the same test, and it has not been independently replicated.
An outside index is plainer
Artificial Analysis, as reported by R&D World, is the clearest third-party counterweight. It places K3.1 Pro at 55 on its Intelligence Index and ranks it fourth of 215 models. Opus 5.5, dated September 22, leads at 58. Claude Sonnet 5.5, dated September 28, and Meta’s Muse Wonder Full, introduced October 12, both score 56. Fable 5.1, GPT-6 Astra, and Google’s Gemini 4 Argon, dated September 30, each score 53. Fourth is a published rank. It does not mean the model caught the frontier. The write-up does not say how Artificial Analysis ordered the tie at 56, so read the scores rather than the place alone.
The cost column is the part a finance lead should read twice. Estimated cost per Intelligence Index task is $2.11 for K3.1 Pro, $5.98 for Opus 5.5, $7.63 for Fable 5.1, and $3.26 for GPT-6 Astra. Cheaper than the models it is chasing. Not the cheapest path to a nearby score: Gemini 4 Argon comes in at $1.99 per task with an index of 53, and GPT-6.1 Sol at $0.72 with a 52.
The reason is volume. K3.1 Pro used 150 million output tokens to finish the index, against a median of 87 million, at 77 tokens per second. A lower sticker, and about 70 percent more output than the median model. Same token price as GPT-6.1 Sol, and a much larger bill once the model keeps talking. In an agent loop, that habit is the invoice.
Peter Gostev, AI capability lead at Arena, formerly LMArena, was blunter about a narrower test. In an October 23 video covering 3D scenes, animations, and websites, he put K3.1 Pro above GPT-5.6 Sol and below GPT-6.1 Sol.
At least in these tests, I don’t think it’s quite there.
One expert, one class of build task. It points the same way as the index: close on some work, not across the board.
What to pilot, and what to leave alone
For coding agents on terminal-style engineering work, a controlled pilot is rational. The near-tie on Terminal-Bench 4.0 and the SWE-Marathon lead are specific enough to try against your own tickets, if you keep them labeled as Moonshot’s numbers until your evals say otherwise. Cap how long a trace can run. Score the runs yourself. Do not assume Highspeed’s two-to-three-times claim will show up on your jobs until you measure it.
Do not move scientific terminal work, or any process where a quiet miss is expensive, on the back of the coding chart. That science gap is the clearest weakness in the public numbers around this release. A coding-agent result does not transfer.
Self-hosting is a later conversation, and a harder one than a date on a calendar. Weights are not public, so you cannot run this yourself today. DeepSeek’s V4.1 Flash line and Zhipu’s GLM-5.3, the open models on the comparison sheet, already can. For K3, Moonshot’s serving guidance pointed at supernodes of 64 or more accelerators, MXFP4 weights, and MXFP8 activations. Separate hardware guidance for K3.1 Pro has not been published. Waiting on November 4 to escape the API bill only works if you already operate that class of cluster.
Pilot the code. Price the extra tokens. Wait for the weights, the license, and your own evals before you call it a platform.
Questions worth settling first
Is K3.1 Pro a new model?
No. It is a further-trained Kimi K3, listed at the same 2.8 trillion total and 104 billion active parameters, with a 1-million-token context. The architecture write-up is the K3 report, not a new design paper.
Can we self-host it on November 4?
Not on the evidence in hand. November 4 is Moonshot’s stated deadline for the weights, and the license has not been disclosed. Until both exist, proprietary is the label procurement should use.
Does the coding chart mean it matches Opus?
Not as an independent finding. Moonshot’s sheet has it 0.7 points behind Opus 5.5 on one terminal-coding test and ahead on another. Your own tickets are the confirmation that matters.
Where is the obvious miss?
Scientific terminal work. On Moonshot’s science bench it sits about 25 points behind Opus 5.5. A good coding-agent run does not carry over.
Is the $2 rate actually cheaper?
Cheaper per token than Opus 5.5, and about $2.11 per Artificial Analysis index task, under Opus, Fable, and Astra. It is not cheaper to run than GPT-6.1 Sol, at $0.72 per task, because this model emits more tokens. Highspeed, at $4 and $20, is no longer the discount tier.
What should a pilot refuse to assume?
That the chart bars have been independently confirmed. They are the vendor’s argument. Also refuse to assume that Highspeed is two to three times faster on your jobs, that temperature can be tuned, or that trimming an agent transcript is free.
What is still unknown on launch day?
The license, whether the weights actually arrive by November 4, which workload Highspeed speeds up, hardware guidance specific to this model, and any independent replication of the headline scores.