Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresVisual Cipher Bench (Einsia.AI, Navers Lab and Tsinghua University) gives coding agents 40 puzzles presented only as images, run in Harbor with a 60-minute limit and no network access; GPT models run in Codex CLI and the others in Claude Code. The score is accuracy over three runs of every puzzle (120 attempts). Most models score at or near zero, so it is shown as a diagnostic outside the capability fit.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6 Astra | max · Codex | 26.7 | |
| 2 | Claude Opus 5.5 | x-high · Claude Code | 21.7 | |
| 3 | GPT-6.1 Sol | max · Codex | 21.7 | |
| 4 | Claude Sonnet 5.5 | low · Claude Code | 0.8 | |
| 5 | GPT-6 Luna | max · Codex | 0.8 | |
| 6 | Kimi K3.1 Pro | max · Claude Code | 0.0 |