Kimi K3.1 Pro: $2/$10 per million tokens, a 55 on Artificial Analysis, and open weights promised by November 4

Update, October 25, 2026: On October 24, VentureBeat reported allegations, based on screenshots posted on X by user Elshayib that it noted haven't been independently authenticated, that creators were offered a Kimi promotional campaign that told them not to use a paid-partnership label; it's unclear who issued the instructions, VentureBeat identified no response from Moonshot, and no public response had appeared by October 26. Separately, Cerebras used Kimi K3 as its worked example for serving trillion-parameter models; see our Cerebras story.

Kimi K3.1 Pro is Moonshot AI's new flagship, released on October 23, 2026, and pitched as an open model even though its weights aren't out yet. Moonshot calls it "our most capable 3T-class open-weight model to date" and says the full weights will be released by November 4, 2026. Until then it's an API and app model, priced at $2 input and $10 output per million tokens, the same list price as OpenAI's GPT-6.1 Sol, which is compared with its siblings in our GPT-6 Astra vs 6.1 Sol vs Luna guide.

This story is based on Moonshot's launch post, its API platform page and quickstart docs, the Kimi K3 model card, Artificial Analysis's launch article and model page, and the Arena and Vals AI leaderboards, checked October 24 and 25, 2026.

The short version

• What it is: a post-trained version of Kimi K3 "with substantially larger post-training compute," with a 1M-token context and image and video input. Thinking is always on.
• Size: Moonshot says "3T-class." Artificial Analysis lists 2.8 trillion total and 104 billion active parameters, the same as Kimi K3's model card.
• Price: $2 input, $0.20 cache hit, $10 output per million tokens. K3.1 Pro Highspeed costs $4 and $20.
• Independent score: 55 on the Artificial Analysis Intelligence Index v4.3.2 (October 23), fourth overall.
• Open weights: promised "by Nov 4, 2026." No license or Hugging Face repository has been named.

How much does Kimi K3.1 Pro cost?

Moonshot's API platform lists two versions. Both have a 1M-token context, and cache writes are billed at the input rate.

Last verified: October 25, 2026

Artificial Analysis says these rates are 33% lower than Kimi K3's on all three lines. Moonshot's launch post says Highspeed delivers "2-3× the speed of K3.1 Pro."

List price isn't the whole cost. Reasoning models differ in how many tokens they spend, as our comparison of cheap, fast models explains. Artificial Analysis says K3.1 Pro used 55,000 output tokens per Intelligence Index task, 14% more than Kimi K3, at about $1.98 per task in its launch article. Its model page showed $2.11 per task on October 25.

The model is also in the Kimi apps, Kimi Work, Kimi Code and Kimi Business, where administrators have to switch it on.

What do Moonshot's own benchmarks show?

All of these are vendor-reported. Moonshot ran them at "max" reasoning effort, temperature 1.0, mostly with the Claude Code harness for coding tasks, and took competitors' scores from their blogs, system cards or leaderboards.

• Terminal-Bench 4.0: 65.7%, against GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%. That's the source of headlines saying K3.1 Pro "tops GPT-6 Astra"; Moonshot's own table puts it narrowly behind Opus 5.5.
• OfficeQA Pro v2: 64.4, which Moonshot calls state of the art.
• Kimi Hallu Bench Hard (Moonshot's own test, lower is better): an 8.3% claim-level hallucination rate with tools, down from 23.4% for K3. Moonshot says these stress-test rates "do not represent overall error rates in daily use."
• Cyber: Moonshot says it removed cyber-related data from training and that K3.1 Pro's cyber capability "remains behind" frontier proprietary models.

What do independent evaluators say?

Each of these evaluators measures something different, so their results aren't combined here.

Ahead of it on the Artificial Analysis index on October 23 were Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Artificial Analysis says its Terminal-Bench 4.0 score is "the highest score of any model" on that test in its runs, ahead of Claude Sonnet 5.5 at 64%. That doesn't match Moonshot's chart, which used Opus 5.5's highest published score; the two used different harnesses. Vals AI's Terminal-Bench 4 board, last updated October 22, puts Claude Opus 5.5 first at 65.15% and doesn't include K3.1 Pro yet.

For context, Artificial Analysis scored Mistral Large 4's preview 38 on the same index, as our Mistral Large 4 story reports.

Is Kimi K3.1 Pro open source?

Not yet. Moonshot's launch post says "Full model weights will be released by Nov 4, 2026." Its docs and launch post don't name a license or a Hugging Face repository. Artificial Analysis's model page still says the weights "are not publicly available," and its launch article says K3.1 Pro "will be the leading open-weights model once its weights are published."

Kimi K3's weights were released under the custom Kimi K3 License, which adds conditions for large commercial users, as our open-model cost comparison describes. Whether K3.1 Pro uses the same license is unconfirmed.

How big is Kimi K3.1 Pro?

Moonshot says only "3T-class." Artificial Analysis lists 2.8 trillion total and 104 billion active parameters, matching Kimi K3's model card: a mixture-of-experts model with 896 experts, 16 active per token. Moonshot hasn't published a K3.1 model card.

What happens to Kimi K2 models?

Moonshot's pricing docs say K2-series models won't be available to new API users from October 31, 2026, and that all K2 models are planned for retirement by the end of December 2026. That includes Kimi K2.6, covered in our Kimi K2.6 story.

What's confirmed, reported and uncertain

• Confirmed (Moonshot): the October 23 release, the 1M context, the API prices, the Highspeed variant, the vendor-reported benchmarks and settings, and the promise of weights by November 4.
• Confirmed (third parties): Artificial Analysis's score of 55 and component results; Arena's preliminary WebDev rank.
• Uncertain: the license, the exact parameter count for K3.1 Pro specifically, and whether the weights arrive on time.