Find the best models on benchmarks

Coding agent and model results on public benchmarks such as Terminal-Bench, run and published by SelfBench users: accuracy and cost per task. By dari.dev.

• terminal-bench/terminal-bench-2-1 87 tasks, 5 settings, run by mupt-ai. Most accurate: Dari Router, 83.9% at $0.11 per task. Best for less: GPT-6 Luna, 65.5% at $0.018.
• datacurve/deep-swe-1-1 113 tasks, 7 settings, run by mupt-ai. Most accurate: GPT-6.1 Sol, 75.2% at $0.83 per task. Best for less: GPT-6.1 Sol, 71.7% at $0.69.