Kimi K3.1 Pro is Kimi's open-weight AI model, released 23/10/2026. Moonshot's model card reports Kimi K3.1 Pro at maximum reasoning effort (temperature 1.0, top-p 0.95) against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. It is close to Claude Opus 5.5 on Terminal-Bench 4.0 (65.7 against 66.4) and ahead on DeepSWE v1.1 and BrowseComp, but behind the closed models on several reasoning and Program Bench rows. Weights are published under a custom licence, and many rows are Moonshot internal benchmarks.
Visit Kimi K3.1 ProAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does Kimi K3.1 Pro score on independent benchmarks?
Kimi K3.1 Pro ranks #11 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 46.5, as measured on 22 November 2026.
Where does Kimi K3.1 Pro rank against other models?
How good is Kimi K3.1 Pro at agentic coding?
How does Kimi K3.1 Pro do on individual tests?
| Evaluation (run by Artificial Analysis) | Kimi K3.1 Pro | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 67.2% | 67.2% |
| FrontierCode 2.0 | 25.5% | 48.8% |
| Humanity's Last Exam | 52.5% | 74.3% |
| GDPval-AA (Elo) | 1861 | 1866 |
How has Kimi K3.1 Pro's score changed over time?
We have tracked Kimi K3.1 Pro on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the Kimi K3.1 Pro system card report?
Moonshot's model card reports Kimi K3.1 Pro at maximum reasoning effort (temperature 1.0, top-p 0.95) against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. It is close to Claude Opus 5.5 on Terminal-Bench 4.0 (65.7 against 66.4) and ahead on DeepSWE v1.1 and BrowseComp, but behind the closed models on several reasoning and Program Bench rows. Weights are published under a custom licence, and many rows are Moonshot internal benchmarks.
These are Moonshot AI's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Sources: Kimi K3.1 Pro (Hugging Face model card) (02/11/2026), Kimi K3.1 Pro License (02/11/2026). Last checked 22/11/2026.
Coding
| Benchmark | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Terminal-Bench 4.0Claude Code harness; comparison scores are the maximum reported values | 65.7% | 57.9% | 58.2% | 66.4% |
| Terminal-Bench-Science 0.1 | 38.1% | 40% | 68.1% | 63.3% |
| SWE Marathon v1.1 | 51.7% | 45.6% | 42.5% | 46.3% |
| BenchCAD w/o tools | 72.8% | 60.6% | n/a | 73% |
| BenchCAD w/ tools | 92.7% | 92.6% | 95.9% | 96.2% |
| Program Bench | 31.5% | 53.5% | 50% | 65% |
| DeepSWE v1.1 | 75.5% | 67.4% | 74.1% | 74.2% |
| PostTrain Bench v1.2 | 36.5% | 44.6% | 41.9% | 43.8% |
| Kimi Code Bench 2.0 (internal)Moonshot internal benchmark | 77.1% | n/a | n/a | n/a |
| Kimi Code Bench 3.0 Long-horizon (internal)Moonshot internal benchmark | 61.2% | 58.8% | 66.5% | 58.5% |
Knowledge & science
| Benchmark | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| HLE-Full w/ tools2,500-question set, with tools | 70% | 65.6% | 57.2% | 67.7% |
| BioMysteryBench | 80.6% | 81.6% | 78.9% | 81.9% |
| HLE-Full2,500-question set, no tools; GLM-5.3 figure is the text-only subset | 51.1% | 57.6% | 53.6% | 59.8% |
| HLE-Diamond1,000-question set; GLM-5.3 figure is the text-only subset | 39.4% | 55.1% | 65.8% | 60.9% |
| GPQA-Diamond | 94.2% | 93.7% | 96.1% | 93.9% |
Agents & tool use
| Benchmark | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| BrowseCompContext compaction at 200K tokens, 500-step maximum, temperature 1.0, top-p 1.0 | 92.1% | n/a | 91.5% | 90.8% |
| Agents' Last Exam (ALL)Runs that bypass stated requirements score zero | 43.2% | n/a | 41.8% | 42.6% |
| Agents' Last Exam (CLI)Runs that bypass stated requirements score zero | 44.2% | n/a | 44.2% | 40.7% |
| Automation Bench-AA | 73.3% | 59.4% | 68.5% | 69.5% |
| GDPval-AA v2.1 | 1846 Elo | 1754 | 1573 | 1866 |
| AA-Briefcase v1.1 | 1678 Elo | 1672 | 1569 | 1803 |
| APEX-Agents | 69.4% | 68.6% | 64.7% | 73.5% |
| OfficeQA Pro v2 | 64.4% | 54.5% | 54.4% | 62.2% |
| OfficeQA Pro | 67.4% | 68.6% | 66.2% | n/a |
| Frontier Finance | 58.6% | 60.8% | 45% | 56.7% |
| Harvey Legal Agent | 23.7% | 19.1% | n/a | n/a |
| HANDBOOK.md | 44.6% | 38.8% | n/a | 41.5% |
| OSWorld V2.1 | 79.6% | 80.7% | 75.8% | 81.8% |
| Business Operation (internal)Moonshot internal benchmark | 37.8% | 30.7% | 32.3% | n/a |
| DECK-Arena (internal)Moonshot internal benchmark | 1456 Elo | n/a | 1423 | 1253 |
| Design Bench (internal)Moonshot internal benchmark | 957 Elo | n/a | 1076 | 927 |
Multimodal
| Benchmark | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| CharXiv (RQ) w/ pythonTool-augmented, averaged over four runs | 93.4% | 92.7% | 92.8% | 93.8% |
| MMMU-Pro w/ pythonTool-augmented, averaged over four runs | 90% | 88.4% | 87.8% | 89.8% |
| GDP.pdf w/ pythonTool-augmented, averaged over five runs | 85.7% | 83.8% | 85.5% | 83.9% |
| ChartoGraphy w/ pythonTool-augmented, averaged over five runs | 90.2% | 88.4% | 91% | 89% |
| BabyVision w/ pythonTool-augmented, averaged over four runs | 95.4% | 95.1% | 96.6% | 97.7% |
| ZeroBench_main w/ pythonpass@5; tool-augmented, averaged over five runs | 63% | 59% | 58% | 61% |
| GDP.pdf | 34.6% | 27.6% | 34.2% | 30.6% |
| CharXiv (RQ) | 89% | 88.8% | 90% | 92.1% |
| MMMU-Pro | 84.1% | 86.3% | 86.9% | 87.7% |
| ZeroBench_mainavg@5 (pass@5 is 39.0); comparison columns show avg@5 | 29.8% | 23% | 43.2% | 35% |
| PerceptionBench | 68.3% | 63% | 77.6% | 72.1% |
Safety & honesty
| Benchmark | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Kimi Hallu Bench Hard w/ tools (internal)Lower is better; Moonshot internal benchmark | 8.3% | 23.3% | 5.5% | 15% |
| Kimi Hallu Bench w/o tools (internal)Lower is better; Moonshot internal benchmark | 5.4% | 11.4% | 4.2% | 6.8% |
Other
| Benchmark | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| ExploitBenchpass@5 maximum-capability score, 300-turn budget | 74.7% | 83% | 100% | 91% |
| ExploitGym | 20.7% | 30.4% | 42.4% | 34.5% |
| xdev-bench-hard (internal)Moonshot internal benchmark | 81% | n/a | n/a | n/a |
What did safety testing find for Kimi K3.1 Pro?
Safety level: No vendor safety framework rating on the model card
- The model card reports cyber capability rows (ExploitBench 74.7, ExploitGym 20.7, both at maximum effort) but no safety framework rating.
- Moonshot reports internal hallucination benchmarks (Kimi Hallu Bench, lower is better): 5.4 without tools and 8.3 on the hard set with tools.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our Kimi K3.1 Pro deep dive.
Kimi Model Timeline
1,048,576 tokens (about 1M) context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
