SelfBench - earendil-works/pi: leaderboard

Coding agents implement changes rebuilt from the merged pull requests of the open-source repository earendil-works/pi: the pull request becomes the instruction and hidden tests written for it decide success. Each row is one model in one harness (Claude Code, Codex or pi) at one reasoning level. Task success rate, higher is better. Source: selfbench.dev.

Metric: Task success (%) on 43 tasks rebuilt from the repository's merged pull requests (hidden tests). Source: selfbench.dev. Saturation forecast: Around February 2027. 10 models tracked. Updated 13 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-6.1 SolHigh88.37#12
2Dari RouterDefault86.05
3Mistral Large 4High83.72#143
4Claude Opus 5.5High83.72#7
5GPT-6 SolHigh83.72#29
6Jev RouterDefault83.72
7Claude Haiku 5.5Low79.07#76
8GLM-5.3High79.07#73
9Kimi K3High76.74#59
10GPT-6 LunaHigh65.12#105

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Example tasks

Add a Cloudflare Workers AI Gateway binding transport to the AI package. Workers with an AI binding in the gateway account should be able to use the gateway without storing a Cloudflare API token just for transport authentication.…

Source

Interactive version: theaggregate.ai/benchmark?slug=selfbench-earendil-works-pi · How It Works · Data refreshed daily, snapshot 2026-11-17.