Moonshot AI has released Kimi K3.1 Pro, a post-trained upgrade of Kimi K3 that puts an open-weight-bound model within a point of Claude Opus 5.5 on a hard agentic coding benchmark. It costs a third less than K3 on the Kimi API, runs about twice as fast, and ships alongside a Highspeed variant for latency-sensitive work.
The model is live today in Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the Kimi API. The full weights are not out yet. Moonshot says they will follow by early November.
It is the biggest AI launch of the past day. The Hacker News thread passed 750 points and 460 comments, a separate thread on its Artificial Analysis score climbed past 100 points, and X posts mentioning Kimi K3.1 topped 14,000 in about a day and a half.
Key points
- Coding: 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%), per Moonshot.
- Independent score: 55 on the Artificial Analysis Intelligence Index, up from 44 for K3, ranking fourth of 215 models tracked.
- Price: $2 per million input tokens and $10 per million output, down from $3 and $15 for K3. Cache hits are $0.20.
- Highspeed:
kimi-k3.1-pro-highspeedcosts $4 / $20 for roughly 2–3× the output speed. - Context: 1 million tokens, text and image input, flat pricing with no long-context surcharge.
- Weights: promised by early November. Until then it is API and app only.
What shipped
Moonshot describes K3.1 Pro as the first model in a new K3.1 series and its most capable “3T-class” model so far. Artificial Analysis lists it at 2.8 trillion parameters, the same size as K3, which is a mixture-of-experts model with about 104 billion parameters active per token. That fits Moonshot’s own description: K3.1 Pro is K3 with much more post-training compute, not a new base model.
Most of the work went into reinforcement learning at scale. Moonshot says it rebuilt its RL framework, long-context inference, sandbox infrastructure and rollout scheduling so it could run joint RL across hundreds of task domains at once, with each increase in compute giving stable gains.
Two API models are available:
kimi-k3.1-pro: $2.00 input, $10.00 output, $2.00 cache write, $0.20 cache hit, per million tokens.kimi-k3.1-pro-highspeed: $4.00 input, $20.00 output, $4.00 cache write, $0.20 cache hit, for the same model at roughly 2–3× the speed.
The API speaks the Chat Completions, Responses and Messages formats, so it works with the OpenAI and Anthropic SDKs by changing the base URL. Moonshot pitches it directly at coding agents including Kimi Code, Codex and Claude Code. Zero Data Retention is available for API customers, and enterprise admins have to switch the model on, since it is off by default for organizations.
The numbers
Moonshot’s headline results, all at maximum reasoning effort:
- Terminal-Bench 4.0: 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Kimi ran its model in the Claude Code harness.
- Kimi Code Bench 3.0 Long-horizon (internal): 61.2%, up from 55.95% for K3.
- BenchCAD with tools: 92.7% on CAD-code reconstruction of 3D objects, behind Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
- OfficeQA Pro v2: 64.4, which Moonshot calls state of the art for financial document analysis.
- Hallucination (internal stress test): claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot notes the test is deliberately adversarial and does not reflect everyday error rates.
The independent check is more useful than the vendor table. Artificial Analysis scores K3.1 Pro at 55 on its Intelligence Index, an 11-point jump over K3, and puts it on the cost-versus-intelligence Pareto line. Its measurements also show where the model is weaker:
- Speed: about 81 output tokens per second on Kimi’s API. That is double K3’s 38.5, but still around the median for its price tier.
- Verbosity: 150 million output tokens to run the index, against a median of 87 million in its class. A cheap per-token price buys less when the model writes nearly twice as much.
- Cost per task: about $2.11 per index task, slightly more than K3’s $2.00, despite the lower list price.
That last point is the one to remember. Per-token price fell by a third, but the cost of getting an answer stayed roughly flat because the model reasons longer.
Cyber capability was cut on purpose
The most unusual section of Moonshot’s post is about security. It says vulnerability research ability rose steadily as post-training scaled, and traced that to cyber-related content in its mid-training data and RL tasks. For the open-weight release it removed that data at both stages.
On its internal exploit benchmark, xdev-bench-hard, a checkpoint trained without those removals solved 90.5% of tasks at pass@3. The released K3.1 Pro solves 81.0%. Moonshot’s stated rule is not to release open weights that would push the frontier of offensive cyber capability, and to raise what it releases gradually over time. It says K3.1 Pro still trails proprietary frontier models that have been public for at least three months.
Several top Hacker News comments pushed back, calling the public model “nerfed” in cybersecurity on purpose, the same complaint often aimed at US labs. Defenders will notice the trade-off: the open release is weaker at exploit work than what the same pipeline can produce.
What it means for builders
- A cheaper Opus-class coding option is on the table. If your agent runs on Claude Code or Codex, K3.1 Pro is a drop-in model to A/B test through the Anthropic- or OpenAI-compatible endpoint. Test it on your own repo tasks, not on the vendor’s charts.
- Budget by task, not by token. The model is verbose. Track cost per completed task, and consider capping reasoning effort for routine work.
- Pick the variant by latency need. Highspeed doubles the price for 2–3× speed. That fits interactive tools; batch jobs and long agent runs can stay on the standard model.
- Use long context with caching. Pricing is flat across the full 1 million tokens, and cache hits drop input cost to $0.20 per million. Kimi defaults to a 5-minute cache; workloads with longer gaps should request the 1-hour TTL to avoid paying for writes again.
- Plan any K2 migration. Kimi’s pricing docs say K2-series models close to new sign-ups at the end of this month and retire by the end of the year.
- Self-hosting is a later decision. At 2.8 trillion parameters, running it yourself is a multi-node job. Most teams will reach it through Kimi’s API or third-party hosts once weights land.
What we could not verify
- Most benchmark figures are Moonshot’s own, several on internal benchmarks, and the coding numbers were run in Claude Code rather than each competitor’s native harness.
- The license for the K3.1 weights is not published yet. K3 shipped under a custom Kimi license that allows commercial use with restrictions.
- Some users on X report hitting usage limits quickly in Kimi’s consumer apps. We have not measured that.
What to watch
- Whether the weights land on schedule, and under what license.
- Independent coding results from builders running real agent workloads, especially against Claude Opus 5.5.
- Third-party API hosts and their pricing once weights are public.
- Further K3.1-series models. Moonshot calls Pro the first in the series.
Key Takeaways
- Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, within a point of Claude Opus 5.5, and 55 on the Artificial Analysis index.
- List price drops to $2 / $10 per million tokens, but high verbosity keeps cost per task close to K3.
- A Highspeed variant at $4 / $20 runs about 2–3× faster.
- Weights are promised by early November, with cyber-related training data deliberately removed from the open release.
Image credit: Kimi
