Up to four models
Gemini 3.8 Flash vs Claude Haiku 5.5
On the index (GPT-4 = 100), Gemini 3.8 Flash scores 175 and Claude Haiku 5.5 174. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
Gemini 3.8 Flash175159–191
WeightsClosed
ReleasedSep 2026
Price in / out$0.82 / $4.13
Context1.05M
Core tests taken7 of 9
Claude Haiku 5.5
174158–191
WeightsClosed
ReleasedOct 2026
Price in / out$0.10 / $0.50
Context1M
Core tests taken7 of 9
Gemini 3.8 Flash and Claude Haiku 5.5 are too close to call (1 point apart, within the uncertainty), ahead on 5 of 7 shared core tests, at 8.2× the price.
Test by testthird-party results
Gemini 3.8 FlashClaude Haiku 5.5Every result as a table
| Test | Gemini 3.8 Flash | Claude Haiku 5.5 |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 95.4% | 89.6% |
| Humanity's Last Exam | 44.5% | 30.1% |
| SimpleQA Verified | 69.7% | 23.8% |
| Maths | ||
| FrontierMath (tiers 1-3) | 68.4% | 75.1% |
| AIME-style maths (OTIS mock) | 98.9% | 98.9% |
| ProofBench v1.1 | 48.0% | 89.0% |
| Chess puzzles | 58.0% | 55.0% |
| FrontierMath (tier 4) | 21.9% | 46.3% |
| Coding | ||
| Terminal-Bench 4.0 | 19.7% | 36.4% |
| Code Migration | 36.5% | 44.6% |
| CyberBench v1.1 | 43.8% | 75.5% |
| IOI | 56.9% | 47.3% |
| ProgramBench | 1.0% | 2.5% |
| Vibe Code Bench 1-100 | 18.8% | 20.4% |
| Vibe Code Bench v1.1 | 78.6% | 90.4% |
| FrontierSWE | 19.6% | 43.8% |
| Agents | ||
| APEX-Agents | 64.3% | 56.1% |
| BioMysteryBench | 62.2% | 70.4% |
| SRE Bench | 11.1% | 5.7% |
| Terminal-Bench-Science | 8.6% | 10.0% |
| DTBench | 95.7% | 95.0% |
| Furniture assembly | 31.7% | 47.5% |
| LMCA | 52.9% | 51.1% |
| Mystery game puzzles | 47.0% | 30.0% |
| Professional work | ||
| Finance Agent (v2) | 61.4% | 54.1% |
| EMB | 72.2% | 59.0% |
| HLAB-Vals-1.1 | 12.1% | 4.2% |
| Legal Research Bench | 38.9% | 43.8% |
| MedCode | 48.1% | 51.5% |
| MedScribe | 84.5% | 89.5% |
| Public Benefits Bench v1.1 | 65.3% | 63.2% |
| SAGE | 35.1% | 45.8% |
| Tax Agent Bench | 32.1% | 31.9% |
| Novel problems | ||
| ARC-AGI-2 | 89.2% | 61.3% |
| ARC-AGI-1 | 98.5% | 89.0% |
| ARC-AGI-3 | 10.4% | 0.5% |
| MysteryMechanism | 36.5% | 35.1% |
Tests at least two of these models have taken. Bold rows are the core tests.
