Updated 2026-10-27
Kimi K3.1 Pro Benchmark Results
Official Moonshot K3.1 Pro benchmark results: Terminal-Bench 4.0, Kimi Code Bench, finance evals, fact accuracy, and safety scores explained.
TL;DR
- Moonshot reports K3.1 Pro at 65.7% on Terminal-Bench 4.0, between Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%.
- On Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scores 61.2%, up from 55.95% for K3.
- The strongest published areas are long-horizon coding, agentic work, financial documents/spreadsheets, 3D workflows, and factual accuracy.
- These are vendor-published results, not independent K3.1 Studio tests; treat them as a shortlist for your own benchmark suite.
How these results should be read
This page summarizes Moonshot AI's official K3.1 Pro documentation and does not claim independent replication. Vendor benchmarks are useful for spotting capability focus, but model ranking can change with prompt format, tools, context length, retries, and scoring rules.
For a production decision, keep at least 20–50 tasks sampled from your real workload, fix the model version ID, use the same tools and timeout, and report pass@1, cost, latency, and number of retries. Replace a model only when the whole bundle improves.
Kimi K3.1 Pro benchmark scores
| Benchmark | Kimi K3.1 Pro | Comparison | What it measures |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Claude Opus 5.5: 66.4%; GPT-6 Astra: 58.2% | Real terminal tasks with strict per-task tests |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | Kimi K3: 55.95% | Multi-step coding agents and end-to-end engineering |
| Office QA Pro v2 | Best reported | Above Opus 5.5, Fable 5.1, GPT-6 Astra | Complex financial-document analysis |
| SpreadsheetBench 2 | Best reported | Above Opus 5.5, Fable 5.1, GPT-6 Astra | End-to-end spreadsheet workflows |
| Fact-statement error rate | 6.85% | K3: 23.35% | Stress test on hard, error-prone conversations |
| xdev-bench-hard | 81.0% | Unfiltered training path: 90.5% | Vulnerability-research capability after safety filtering |
Where K3.1 Pro wins
- Long coding sessions: Moonshot emphasizes fixes in real repositories, project creation, and coordination of terminal tools.
- Knowledge and finance work: official evals cover cross-document checking, spreadsheet formulas, financial modeling, and investment workflows.
- 3D and spatial work: the model can write and run Blender, SolidWorks, Rhino, Unity, Unreal, and Godot scripts without an MCP connector, according to Moonshot.
- Repeated long prompts: 1M context plus automatic prefix caching can lower cost when the same knowledge base is reused.
Where caution is required
- Terminal-Bench 4.0 still favors Claude Opus 5.5 by 0.7 percentage points in Moonshot's own table.
- Finance and code benchmarks include internal suites that are harder for third parties to reproduce identically.
- Vision input does not accept public image URLs directly; use base64 or an ms:// file ID.
- Moonshot says web search is being updated and is not currently recommended for production flows.
Frequently asked questions
How good is Kimi K3.1 Pro at coding?
Moonshot reports 65.7% on Terminal-Bench 4.0 and 61.2% on Kimi Code Bench 3.0 Long-horizon. The latter is 5.25 percentage points higher than Kimi K3.
Is Kimi K3.1 Pro better than Claude?
On Moonshot's Terminal-Bench 4.0 table, Claude Opus 5.5 scores 66.4%, slightly ahead of K3.1 Pro at 65.7%. Your own coding-agent benchmark should decide the tie.
Are these benchmarks official or independent?
Official. They come from Moonshot AI's K3.1 Pro documentation. K3.1 Studio has not independently reproduced them.