Find the best models for your repo

SelfBench measures AI coding agents and models on tasks from a repository’s own merged pull requests, to show which works best on your code. By dari.dev.

• Next.js Framework Work 2 repositories (dubinc/dub.link, calcom/cal.diy), 56 tasks, 12 settings, run by mupt-ai.
• earendil-works/pi 43 tasks, 18 settings, run by mupt-ai. Most accurate: GPT-6.1 Sol, 88.4% at $0.53 per task. Best for less: Dari Router, 86.0% at $0.40.
• mupt-ai/self-bench 38 tasks, 7 settings, run by mupt-ai. Most accurate: GPT-6.1 Sol, 76.3% at $0.43 per task. Best for less: DeepSeek V4.1 Flash 1017, 71.1% at $0.13.
• vercel/next.js 38 tasks, 13 settings, run by mupt-ai. Most accurate: GPT-6.1 Sol, 84.2% at $0.91 per task. Best for less: GPT-6.1 Sol, 78.9% at $0.56.
• PostHog/posthog 35 tasks, 14 settings, run by mupt-ai. Most accurate: Claude Opus 5.5, 68.6% at $2.33 per task. Best for less: DeepSeek V4.1 Flash 1017, 65.7% at $0.078.
• getsentry/sentry 78 tasks, 15 settings, run by mupt-ai. Most accurate: Claude Sonnet 5.5, 82.1% at $0.59 per task. Best for less: GPT-6.1 Sol, 80.8% at $0.50.
• supabase/supabase 36 tasks, 12 settings, run by mupt-ai. Most accurate: GPT-6.1 Sol, 77.8% at $0.46 per task. Best for less: GPT-6.1 Sol, 75% at $0.24.
• vercel/eve 38 tasks, 10 settings, run by mupt-ai. Most accurate: GPT-6.1 Sol, 73.7% at $0.43 per task. Best for less: GPT-6 Luna, 50% at $0.040.
• vercel/vercel 47 tasks, 8 settings, run by mupt-ai. Most accurate: Claude Opus 5.5, 80.9% at $1.27 per task. Best for less: Kimi K3, 78.7% at $0.96.
• pingdotgg/t3code 39 tasks, 9 settings, run by mupt-ai. Most accurate: Claude Opus 5.5, 76.9% at $2.05 per task. Best for less: GPT-6.1 Sol, 66.7% at $0.31.