Moonshot AI·released 2026-01-271 source
Kimi K2.5 benchmark scores: 17 benchmarks tracked. 35% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 17 benchmarks·best result 92.19 on OTIS Mock AIME 2024-2025 (reproduced)·6 of 17 independently reproduced·$0.6/$3 per M tokens
Consensus: LiteLLM · Cross-check: models.dev, OpenRouter · See every model’s pricing →
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| OTIS Mock AIME 2024-2025 | 92.19accuracy (%) | reproduced· optimizedT1 | 2026-01-27Epoch AI |
| Fiction.LiveBench | 86.1accuracy (%) | unverifiedT1 | 2026-01-27Fiction.live leaderboard |
| GPQA diamond | 83.47accuracy (%) | reproduced· optimizedT1 | 2026-01-27Epoch AI |
| SWE-Bench verified | 73.76accuracy (%) | reproduced· optimizedT1 | 2026-01-27Epoch AI |
| ARC-AGI | 65.33accuracy (%) | unverified· optimizedT2 | 2026-01-27 |
| OSWorld | 63.3accuracy (%) | unverified· optimizedT1 | 2026-01-27OS World Website |
| FrontierMath-2025-02-28-Private | 48.95accuracy (%) | reproduced· optimizedT1 | 2026-01-27Epoch AI |
| WeirdML | 45.6accuracy (%) | unverifiedT1 | 2026-01-27https://htihle.github.io/weirdml.html |
| Terminal Bench | 43.2accuracy (%) | unverifiedT1 | 2026-01-27https://www.tbench.ai/leaderboard/terminal-bench/2.0 |
| SimpleBench | 36.16accuracy (%) | unverifiedT1 | 2026-01-27SimpleBench Leaderboard |
| SimpleQA Verified | 34.3accuracy (%) | unverifiedT2 | 2026-01-27 |
| HLE | 20.56accuracy (%) | unverifiedT2 | 2026-01-27 |
| CL-bench | 19.3accuracy (%) | unverifiedT2 | 2026-01-27 |
| CL-bench Life | 13.2accuracy (%) | unverifiedT2 | 2026-01-27 |
| ARC-AGI-2 | 11.81accuracy (%) | unverifiedT2 | 2026-01-27 |
| Chess Puzzles | 7.41accuracy (%) | reproducedT1 | 2026-01-27Epoch AI |
| FrontierMath-Tier-4-2025-07-01-Private | 7accuracy (%) | reproducedT1 | 2026-01-27Epoch AI |
Claims drawn from cited facts, not live model generation.
This model was developed by Moonshot AI and released in 2026. A 2026 vintage places it in the most recent generation of models, so it reflects contemporary architecture and training practices rather than older, superseded designs.
3 cited facts
The model is scored on 17 tracked benchmarks, and it holds no current top score on any of them. Its benchmark results are independently reproduced, meaning outside evaluation backs the scores up.
3 cited facts
Its mean gap to the best recorded score across its tracked benchmarks is 29.25 points.
1 cited fact
6 facts cross-checked across data sources: 1 corroborated, 1 single-source, 4 disputed.
Source facts, citations, and refresh stamp for this record.
Sources: curated_capability_claims, epoch_benchmarks, litellm_prices, modelsdev_models +1 more
Kimi K2.5 is an AI model developed by Moonshot AI, released 2026-01-27. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
Kimi K2.5 has recorded scores on 17 benchmarks, each shown with its evidence status.
Kimi K2.5 has recorded scores on 17 benchmarks — GPQA diamond, OTIS Mock AIME 2024-2025, WeirdML, SimpleQA Verified, Chess Puzzles, SimpleBench, and 11 more. The full table above shows each score with its evidence status.
6 of 17 recorded scores (35%) are independently reproduced rather than self-reported by the lab.
Kimi K2.5 has 17 tracked claims: 6 independently reproduced, 11 unverified.
Listed API pricing: $0.6 per million input tokens, $3 per million output tokens (prices disputed across sources). See the pricing block for the full breakdown.