Atlas

Benchmarks

← All benchmarks

Continual Learning Exam (Trial 2, with feedback)

Agents · 2026-10-25

The Continual Learning Exam's second complete run, which allows one resubmission per level after diagnostic feedback on a failed first submission. Same 120 tasks and 600 levels, grading and serial unlocking as Trial 1; the score is the share of the 600 levels cleared, in percent.

Top models (higher is better)

ModelScore
Claude Opus 547.7
Sonnet 540.2
Gemini 3.7 Flash33.7
GPT-5.6 Sol33.2
GPT-5.6 Terra33.2
Gemini 3.8 Flash30.7
DeepSeek V4 Pro 081327.3
Hy327.3
Qwen3.8-Max26.7
GPT-5.6 Luna25.3
GLM-5.324.0
Gemini 3.6 Flash23.5
Loading Atlas data…