Atlas

Benchmarks

← All benchmarks

Terminal-Bench 4.0 — Science

Agents · 2026-09-22

Resolution rate on the 14 science tasks in Terminal-Bench 4.0.

Top models (higher is better)

ModelScore
Kimi K3.1 Pro73.8
Gemini 4 Carbon59.5
GPT-6 Astra57.1
Muse Wonder Full57.1
GPT-6.1 Sol54.8
GPT-6.1 Terra54.8
Gemini 4 Argon52.4
Muse Wonder47.6
GPT-5.6 Sol42.9
GPT-6 Sol42.9
Muse Spark 1.340.5
Opus 4.838.1
DeepSeek V4.1 Flash 101738.1
GLM-5.338.1
GLM-5.3 Flash33.3
MiMo-V2.6-Pro33.3
Step 5 Preview33.3
Grok 4.728.6
DeepSeek-V4-Flash-073126.2
GPT-5.6 Terra26.2
MiniMax M3.1 Flash26.2
Qwen3.8 Max (0902)26.2
Claude Haiku 5.521.4
DeepSeek V4.1 Flash21.4
Gemini 3.8 Flash21.4
GPT-6 Luna21.4
MiMo-V2.6-Flash21.4
DeepSeek V4 Pro 081319.0
Ember-116.7
Grok 4.616.7
Mistral Large 4 Preview16.7
GPT-5.6 Luna14.3
Grok 4.514.3
Sonnet 511.9
Gemini 3.5 Flash11.9
Gemini 3.7 Flash11.9
Muse Spark 1.211.9
Kimi K39.5
Muse Spark Open7.1
Gemini 3.6 Flash7.1
Loading Atlas data…