Moonshot AI released Kimi K3.1 Pro on October 23, a post-trained update of its July model Kimi K3, and independent tracker Artificial Analysis now ranks it fourth of 215 models on its Intelligence Index with a score of 55. Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) score higher, according to figures reported by The Decoder. The release matters because Moonshot says the full weights will be published by November 4, which would make it the highest-scoring open-weight model on that index.
What Moonshot released
K3.1 Pro keeps the architecture of Kimi K3: 2.8 trillion total parameters, 104 billion active per token, a one-million-token context window and text-plus-image input. The difference is a much larger amount of reinforcement-learning post-training across what the company calls hundreds of downstream domains. Moonshot also shipped a Highspeed variant that it says runs two to three times faster. Both are available on Kimi.com, Kimi Work, Kimi Code and the Kimi API.
What is independently measured
Artificial Analysis lists the model at 55 on its Intelligence Index, an 11-point gain over K3, and notes it is slower than average at about 73 output tokens per second and very verbose: it produced 150 million tokens to complete the index, against a median of 87 million. Its page lists a cost of about $2.11 per index task, while The Decoder cited $1.98 at the time of its report. The tracker also reports a score of 1,846 Elo on its GDPval-AA agentic-work test, behind Opus 5.5 at 1,866.
What Moonshot reports itself
The company’s own benchmark table, which uses maximum reasoning effort, should be read as vendor-reported. It lists Terminal-Bench 4.0 at 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%); HLE-Full with tools at 70.0%; and BrowseComp at 92.1%. On the CAD benchmark BenchCAD it reports 92.7% with tools, below Opus 5.5 at 96.2%. On ExploitBench, an offensive-security test, it reports 74.7% against 91.0% for Opus 5.5 and 100% for GPT-6 Astra. Moonshot also says claim-level hallucinations on its internal “Hallu Bench Hard” fell from 23.4% to 8.3%, and describes that as a stress test, not a general error rate.
Pricing and cyber limits
Moonshot cut API prices by a third: $2 per million input tokens (from $3), $10 per million output tokens (from $15) and $0.20 for cached input. The Highspeed variant costs $4 and $20. Because Artificial Analysis says the model uses about 14% more output tokens, cost per task stays nearly flat against K3.
The company devotes a section of its post to cybersecurity. It says vulnerability-research ability kept rising as post-training scaled, so for the open-weight release it removed cyber-related data from both training phases. On its internal xdev-bench-hard, the uncleaned model solved 90.5% of tasks and the cleaned K3.1 Pro 81.0%. These are Moonshot’s numbers and we have not seen an independent replication.
Why it matters
If the weights ship on schedule, the gap between open and proprietary models on this index would shrink from 12 points to 3, assuming nothing else changes. That is a claim about one index; rankings vary by test, and the benchmark table above shows K3.1 Pro trailing on exploit and CAD tasks. For teams weighing open against closed models, our comparison of open-weight and closed models covers the trade-offs, and the model also lands as Washington restricts a rival Chinese lab, see the DeepSeek federal ban.
What to watch next
- Whether the weights are actually published by November 4, and under what license.
- Independent replications of the agentic and coding scores outside Moonshot’s own harness.
- How Artificial Analysis scores change once third-party providers host the model.
- Responses from Anthropic, OpenAI and DeepSeek on price.
- Whether the cleaned cyber data holds up against fine-tuning by outside developers.
Key takeaways
- Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall.
- Moonshot’s benchmark table is vendor-reported; the index ranking is independent.
- API prices fell 33% while the model size stayed at 2.8 trillion parameters.
- Open weights are promised by November 4, not yet released.
- The model trails Claude Opus 5.5 on exploit and CAD tests in Moonshot’s own table.
Frequently asked questions
Is Kimi K3.1 Pro open-weight today?
Not yet. Moonshot says the full weights will be released by November 4, 2026. Today it is available through Moonshot’s apps and API.
Who ranks Kimi K3.1 Pro fourth?
Artificial Analysis, an independent benchmarking site, on its Intelligence Index. The Artificial Analysis page lists it at 55 points, fourth of 215 models.
How does it compare on price?
Input costs $2 and output $10 per million tokens, down from $3 and $15 for Kimi K3, according to Moonshot and The Decoder.
