Up to four models
Claude Haiku 5.5 vs Kimi K3
On the index (GPT-4 = 100), Claude Haiku 5.5 scores 174 and Kimi K3 173. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
Claude Haiku 5.5
174158–191
WeightsClosed
ReleasedOct 2026
Price in / out$0.10 / $0.50
Context1M
Core tests taken7 of 9
Kimi K3
173156–189
WeightsOpen
ReleasedJul 2026
Price in / out$0.20 / $20
Context1.05M
Core tests taken6 of 9
Claude Haiku 5.5 and Kimi K3 are too close to call (2 points apart, within the uncertainty), ahead on 5 of 6 shared core tests, and costs 26× less.
Test by testthird-party results
Claude Haiku 5.5Kimi K3
Every result as a table
| Test | Claude Haiku 5.5 | Kimi K3 |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 89.6% | 93.1% |
| LiveBench Data Analysis | 72.8% | 78.7% |
| LiveBench IF | 66.5% | 71.4% |
| LiveBench Language | 63.5% | 85.5% |
| LiveBench Reasoning | 84.6% | 90.7% |
| SimpleQA Verified | 23.8% | 50.6% |
| Maths | ||
| FrontierMath (tiers 1-3) | 75.1% | 72.2% |
| AIME-style maths (OTIS mock) | 98.9% | 97.2% |
| LiveBench Mathematics | 92.8% | 84.4% |
| ProofBench v1.1 | 89.0% | 87.0% |
| Chess puzzles | 55.0% | 37.0% |
| FrontierMath (tier 4) | 46.3% | 39.0% |
| Coding | ||
| Terminal-Bench 4.0 | 36.4% | 17.2% |
| LiveBench Agentic Coding | 57.2% | 62.2% |
| LiveBench Coding | 77.9% | 81.5% |
| Code Migration | 44.6% | 16.1% |
| CyberBench v1.1 | 75.5% | 75.2% |
| IOI | 47.3% | 48.9% |
| ProgramBench | 2.5% | 2.0% |
| Vibe Code Bench 1-100 | 20.4% | 18.2% |
| Vibe Code Bench v1.1 | 90.4% | 85.0% |
| FrontierSWE | 43.8% | 25.9% |
| Agents | ||
| APEX-Agents | 56.1% | 47.8% |
| BioMysteryBench | 70.4% | 72.6% |
| Terminal-Bench-Science | 10.0% | 1.4% |
| DTBench | 95.0% | 91.2% |
| Furniture assembly | 47.5% | 34.2% |
| LMCA | 51.1% | 52.7% |
| Mystery game puzzles | 30.0% | 26.0% |
| Professional work | ||
| Finance Agent (v2) | 54.1% | 53.1% |
| EMB | 59.0% | 66.7% |
| HLAB-Vals-1.1 | 4.2% | 12.9% |
| Legal Research Bench | 43.8% | 46.2% |
| MedCode | 51.5% | 49.4% |
| MedScribe | 89.5% | 88.0% |
| Public Benefits Bench v1.1 | 63.2% | 68.2% |
| SAGE | 45.8% | 52.8% |
| Tax Agent Bench | 31.9% | 33.4% |
| Novel problems | ||
| ARC-AGI-2 | 61.3% | 60.4% |
| ARC-AGI-1 | 89.0% | 94.5% |
Tests at least two of these models have taken. Bold rows are the core tests.
