Moonshot AI released Kimi K3.1 Pro on October 23, 2026, and the two headline numbers look like the same story when they are actually two different stories. API prices dropped 33 percent: input tokens now cost $2 per million (down from $3), output tokens $10 (down from $15), cached input $0.20 (down from $0.30). Meanwhile the model jumped 11 points on the Artificial Analysis Intelligence Index, to 55, good for fourth place among 215 models and only 3 points behind Claude Opus 5.5 at 58, as The Decoder and R&D World both report.
The catch, and the most useful number in the whole release: the price cut bought users almost nothing in practice. Artificial Analysis, working from token usage and list prices, calculated that a K3.1 Pro model generates roughly 14 percent more output tokens than K3, which lands the cost per Intelligence Index task at about $1.98, against $2.00 for the older model. Cheaper tokens, same bill, because the model writes more of them.
Context: Same Model, Far More Post-Training Compute
Kimi K3.1 Pro is not a new pre-trained model. Moonshot describes it in its technical blog as a post-trained version of Kimi K3 with "substantially larger post-training compute," keeping the same dimensions: 2.8 trillion total parameters, 104 billion active per token (a sparse mixture-of-experts design), a one-million-token context window, and text plus image input. Full model weights are promised by November 4, 2026. A companion variant, K3.1 Pro Highspeed, runs at 2 to 3 times the speed and costs twice as much per token ($4 in, $20 out).
That lineage is what makes the release interesting to anyone who trains models. Moonshot optimized "every layer of the stack" for its reinforcement learning runs, including the RL framework, long-context inference, sandbox infrastructure, and rollout scheduling, so that each increment of RL compute across hundreds of downstream domains delivered stable returns. The result is close to a controlled ablation of post-training compute: identical base weights, much more RL, and benchmark scores that move by double digits.
| Benchmark | K3.1 Pro | K3 | Best rival (source: Moonshot sheet unless noted) |
|---|---|---|---|
| AA Intelligence Index (independent) | 55 | 44 | Claude Opus 5.5: 58 |
| Terminal-Bench 4.0 | 65.7% | 13% | Claude Opus 5.5: 66.4%, GPT-6 Astra: 58.2% |
| GDPval-AA v2.1, Elo (independent) | 1,846 | 1,534 | Claude Opus 5.5: 1,866 |
| SWE Marathon v1.1 | 51.7 | n/a | Claude Opus 5.5: 46.3 |
| HLE-Full with tools | 70.0 | n/a | Claude Opus 5.5: 67.7 |
| Kimi Hallu Bench Hard, claim-level hallucination | 8.3% | 23.4% | GPT-6 Astra: 5.5% (per The Decoder) |
Moonshot reports top scores on six of its 14-test comparison sheet, including SWE Marathon (51.7 vs 46.3 for Opus 5.5) and OfficeQA Pro v2 (64.4 vs 62.2), a financial-document analysis test. The 13 percent to 65.7 percent move on Terminal-Bench 4.0 is the single loudest figure in the release: agentic terminal task pass rates that were unimpressive in K3 became frontier-competitive in K3.1 Pro without touching the pre-trained weights.
The Benchmark Sheet Deserves Its Footnotes
R&D World's methodology note is worth more attention than most benchmark tables get. Moonshot ran K3.1 Pro at maximum reasoning effort (temperature 1.0, top-p 0.95), mostly on the Claude Code harness, and its rival scores mix Moonshot's own runs with vendor system cards and public leaderboards. Kimi Code Bench 3.0 and Kimi Hallu Bench are internal benchmarks. The DeepSeek score on Terminal-Bench Science comes from an older DeepSeek version; Moonshot's own run of the current one scored 9.3.
And the sheet has genuine weak spots. On Terminal-Bench Science 0.1, which tests scientific research workflows, K3.1 Pro scores 38.1 against 68.1 for GPT-6 Astra and 63.3 for Opus 5.5. On ExploitBench it posts 74.7 while GPT-6 Astra reaches 100 and Opus 5.5 hits 91.0, though as we will see, at least part of that gap is deliberate.
Independent signals are broadly positive but not breathless. Artificial Analysis places K3.1 Pro at the top of its Pareto frontier for its parameter class: no comparably sized model scores higher per dollar. Peter Gostev, AI capability lead at Arena (formerly LMArena), tested the model on 3D scenes, animations and websites and told viewers in an October 23 video that its work ranked above GPT-5.6 Sol's but below GPT-6.1 Sol's. "At least in these tests, I don't think it's quite there," he said, regarding the frontier.
Why This Matters: Three Numbers Engineers Should Actually Budget On
The price cut is a pricing-page event, not a budget event. This is the third consecutive launch where list price and effective cost told different stories, and it is becoming the pattern we track on this site: GPT-6.1 Sol's flat price against falling cost per task and OpenAI's permanent pricing change both taught the same lesson. Buyers who re-forecast spend from "$10 instead of $15 per million output tokens" will underspend the forecast by roughly the verbosity delta. Artificial Analysis counted 150 million output tokens from K3.1 Pro to complete its index against a median of 87 million, with output speed measured at 77 tokens per second. Longer reasoning and more tool calls are what buy the benchmark gains; the invoice has to carry them.
That said, the cross-model economics are genuinely strong. R&D World reports Artificial Analysis's weighted average cost per Intelligence Index task: $2.11 for K3.1 Pro, versus $5.98 for Claude Opus 5.5, $7.63 for Claude Fable 5.1, $3.26 for GPT-6 Astra, $1.99 for Gemini 4 Argon, and $0.72 for GPT-6.1 Sol. The list-price overlap with the US frontier is new territory: $2/$10 matches GPT-6.1 Sol and is half of Opus 5.5's rate, a convergence we flagged when Meta's Muse Spark Open shipped, when Chinese and Western API pricing first started to blur.
Post-training compute is now the purchasable frontier. An 11-point index gain and a 5x improvement in Terminal-Bench pass rate from the same base weights is the clearest public demonstration yet that, at this performance level, capability is bought after pre-training. For teams running their own fine-tuning or RL pipelines, the read is that Moonshot's stack work (rollout scheduling, sandbox infrastructure, environment quality checks) is the moat, not the checkpoint. This echoes what DeepSeek's DeepInfer disclosures showed on the serving side: the Chinese labs are increasingly publishing the engineering, not just the weights.
The cyber ablation is the most honest safety table in open-weights history. Moonshot traced its model's growing vulnerability-research capability to cyber-related content embedded in the general mid-training corpus and RL environment mix, removed that data from both stages for the open-weights release, and then published the negative control: on its internal xdev-bench-hard (recently disclosed vulnerabilities in the Linux kernel, Node.js, JavaScriptCore, Exim, with mitigations on), the uncleaned checkpoint solves 90.5 percent of tasks at pass@3, the shipped model 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on the same test and K3 scored 33.3 percent. Two things follow. Data removal buys roughly 9.5 points of measured capability suppression, which is a real, quantified safety lever rather than a vibe. And 81.0 percent still beats most closed models, so the "we removed the cyber data" label should not be read as a capability cliff. Moonshot's stated policy is to keep open-weight cyber capability below the absolute proprietary frontier while gradually raising it for defensive work, auditing legacy code on infrastructure users control themselves.
Outlook
If the weights land by November 4 as promised, K3.1 Pro would be the highest-scoring open-weights model available, cutting the open-versus-proprietary gap on the Intelligence Index from 12 points to 3, per The Decoder's framing. The self-hosting math is the part nobody can run yet: 2.8 trillion total parameters, even at 104 billion active, is a multi-node deployment on current hardware, and the memory-versus-compute split that defined Muse Spark Open's launch applies here at roughly three times the footprint.
For teams already routing between OpenRouter-style gateways, the practical takeaway is narrower than the headlines. K3.1 Pro is now a credible default for long-horizon coding and financial-document work at half of Opus 5.5's list price, while research-workflow agents and anything needing maximum exploit-grade analysis still belong on the closed frontier models, at least on the evidence published this week.
Three things to watch. First, third-party harness results: Moonshot's best agentic scores came partly from running on Anthropic's Claude Code harness, so expect numbers to move once teams rerun on their own scaffolds. Second, whether the 14 percent verbosity tax holds under cheaper decoding settings, because that single multiplier decides whether the 33 percent price cut ever reaches a bill. Third, the November 4 weights themselves, including whether the cyber data removal survives inspection the way its benchmarks suggest. A 3-point gap to the frontier is small enough that the next two releases, from Moonshot or from Anthropic, decide who closes it.