Project APE CRED: leaderboard

Models detect a planted error in a supplied solution. Detection recall uses 200 trials for each API or agent configuration, with agent harnesses and versions retained separately.

Metric: Detection recall (%). Source: ape.socialcatalystlab.org. 24 models tracked.

Top models

#ModelSettingScoreOverall rank
1GPT-6 Astra(Low) (APE CRED run: harness=Codex CLI, version=0.153.4, native=codex6astra_low)99.5#8
2GPT-5.6 Sol(High) (APE CRED run: harness=Codex CLI, version=0.144.4, native=codex56sol_high)99#33
3GPT-5.6 LunaxHigh97.5#130
4GPT-5.6 Terra(High) (APE CRED run: harness=Codex CLI, version=0.144.4, native=codex56terra_high)97#68
5GPT-5.5(High) (APE CRED run: harness=Codex CLI, version=0.142.5, native=codex_high)96.5#50
6Claude Fable 5(High) (APE CRED run: harness=Claude Code, version=2.1.215, native=fable5_high)96.5#28
7GPT-5.6 Pro Luna91#98
8GLM-5.3 FlashxHigh90#136
9GLM-5.2High89.5#152
10GPT-6 Luna(High) (APE CRED run: harness=Codex CLI, version=0.156.1, native=codex6luna_high)87.5#105
11Claude Opus 4.8(Medium) (APE CRED run: harness=Claude Code, version=2.1.118, native=opus48_med)85#62
12DeepSeek V4 ProxHigh83.5#198
13Claude Sonnet 4.6(High) (APE CRED run: harness=Claude Code, version=2.1.107, native=sonnet_high)83.5#173
14DeepSeek V4 FlashxHigh80.5#265
15Gemini 3.8 FlashHigh79.5#39

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

Interactive version: theaggregate.ai/benchmark?slug=project-ape-cred · How It Works · Data refreshed daily, snapshot 2026-11-17.