Kimi K3.1 Pro is here: Moonshot's 2.8T model nearly matches Claude Opus 5.5 on coding, open weights due November 4

• Moonshot AI released Kimi K3.1 Pro on October 23, 2026 — the first model in the K3.1 series, announced on X and in a technical blog post.
• Specs: 2.8 trillion total parameters with 104 billion active in a mixture-of-experts design, a 1-million-token context window, and text+image multimodal input. It is a post-trained Kimi K3, not a new base model.
• Moonshot's own numbers: 65.7% on Terminal-Bench 4.0 versus 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5 — company-reported and not independently verified.
• Independent check: Artificial Analysis gives it 55 on its Intelligence Index, fourth of 215 models — it would be the top open-weight model once the weights ship.
• The full open weights are promised by November 4, 2026; the API is live now at $2/$10 per million input/output tokens, 33% below K3's prices.
• For the open release Moonshot stripped cyberattack data from training: an unfiltered checkpoint solved 90.5% of its internal xdev-bench-hard exploit tasks, the released model 81.0%.

Moonshot AI released Kimi K3.1 Pro on Friday, calling it its most capable 3T-class open-weight model to date. It's the first of the K3.1 series: Kimi K3 put through far heavier post-training, and by the company's own numbers it lands within a point of Anthropic's Claude Opus 5.5 on one of the toughest coding-agent tests in the business. The catch is familiar — the most eye-catching numbers are Moonshot's own — but the independent data looks strong too, and the full weights land by November 4.

Within striking distance of the closed frontier

On Terminal-Bench 4.0, which grades coding agents on real terminal tasks with strict per-task tests, Moonshot reports 65.7% for K3.1 Pro against 58.2% for OpenAI's GPT-6 Astra and 66.4% for Claude Opus 5.5 — less than a point off the Anthropic flagship. Moonshot says K3.1 Pro ran inside the Claude Code harness for most coding tests, while rival scores came from public leaderboards or the rivals' own reports. It also claims a top score of 64.4 on OfficeQA Pro v2, a test of analysis across real financial documents, and 92.7% with tools on BenchCAD, behind Opus 5.5's 96.2% and GPT-6 Astra's 95.9%.

The independent yardstick is kinder: Artificial Analysis puts K3.1 Pro at 55 on its Intelligence Index, fourth of 215 models, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Meta's Muse Wonder Full (56). If the weights ship on schedule, it becomes the highest-scoring open-weight model on the index — cutting the gap between the best open and closed models from 12 points to 3. Other gains Moonshot cites versus K3: GDPval-AA up 312 points to 1,846 Elo, AutomationBench-AA from 58% to 73%, Humanity's Last Exam from 47% to 52%.

A post-training play, not a new brain

K3.1 Pro is not a new base model. Moonshot describes it as a post-trained Kimi K3, built with substantially larger post-training compute and joint reinforcement learning across hundreds of downstream domains. The architecture keeps K3's size: 2.8 trillion total parameters with 104 billion active in a mixture-of-experts design, a 1-million-token context window, and text-plus-image input.

Alongside it comes Kimi K3.1 Pro Highspeed, which Moonshot says runs two to three times faster at the same capability level. Nathan Lambert, who writes the Interconnects newsletter on open models, wrote the scores "look to be clearly the strongest current open weight model, if the weights are released as they claim by November 4th" — adding that Moonshot seems to have "figured out their RL, where the initial release was a very impressive base model with lighter post-training."

Cheaper to run, live today

Both models are live now in the Kimi apps, the Kimi Work desktop app (from version 3.1.0), and Kimi Code, and through the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with zero data retention available to API customers. Kimi Enterprise admins have to opt in — access is off by default. The full model weights are promised by November 4.

Pricing is aggressive: input costs $2 per million tokens, output $10, cached input $0.20 — a 33% cut versus K3. The Highspeed model costs $4 and $20. On an internal stress test, Moonshot says the claim-level hallucination rate fell from 23.4% on K3 to 8.3%, with its own caveat: "The reported rates do not represent overall error rates in daily use." Existing Kimi users should note the K2 series is being retired: new sign-ups lose access October 31 and the series shuts down by the end of December.

The cyber line Moonshot drew

The most interesting part of the launch is what Moonshot deliberately left out. As post-training scaled up, the company's vulnerability-research scores kept rising — and engineers traced the cause to cyber content mixed into the general mid-training corpus and cyber tasks in the RL environments. For the open-weight release, Moonshot removed both and kept the general coding tasks.

The yardstick is xdev-bench-hard, an internal suite that recreates recently disclosed bugs in software like the Linux kernel and Node.js and asks the agent to write a working exploit. The checkpoint trained without filtering hit 90.5% at pass@3; the released model scores 81.0% — the 9.5-point gap is Moonshot's own measure of what the removal took away. Even so, 81% is well above K3's 33.3%. The company's stated principle: "we aim not to release open weights that would extend the absolute frontier of cyber capability." It also says it intends to raise the cyber capability of its open weights over time so more teams can audit the code they maintain on their own infrastructure.

The announcement lands in a heated debate. Anthropic published a report in late September arguing that an open model from China's Z.ai posed a cyberattack risk precisely because anyone can download and modify it; on October 14, OpenAI CFO Sarah Friar told Bloomberg Television, "We think you want to be on more Western models on the cyber front." One honest gap: Moonshot's blog post doesn't address whether someone could fine-tune the released weights to put the filtered capability back. The November 4 weights drop will let outside researchers check both the benchmark claims and whether the cyber filtering holds up.