Haiku-tier model matching Sonnet 4 on coding at one-third the cost
| Benchmark | Score | Rank |
|---|---|---|
OSWorld Computer use in real desktop environments | 50.7% | #9 / 9 |
MATH Competition-level mathematics problems | 95.3% | #16 / 50 |
MMLU Tests knowledge across 57 subjects from STEM to humanities | 90.8% | #18 / 54 |
SWE-bench Real-world GitHub issue resolution | 73.3% | #20 / 40 |
ARC-AGIARC Prize Novel reasoning tasks requiring fluid intelligence | 5.1% | #21 / 23 |
Terminal Agentic terminal coding tasks requiring multi-step execution | 41% | #22 / 48 |
MMLU-Provals.ai Harder 10-option successor to MMLU; more reasoning-focused | 78.7% | #35 / 38 |
HumanEval Coding ability - generating correct Python functions | 85.2% | #37 / 50 |
LiveCodeBenchvals.ai Contamination-free competitive programming (filtered by cutoff date) | 41.2% | #38 / 40 |
MMMUvals.ai College-level multimodal reasoning across 30+ disciplines | 46.1% | #39 / 39 |
GPQA PhD-level science questions even experts struggle with | 73% | #46 / 64 |