Atlas

Benchmarks

← All benchmarks

ProgramBench Average Hidden-Test Pass Rate

Code

Mean hidden-test pass percentage across the 200 ProgramBench tasks, including partial credit. Vals calls this Raw Pass Rate. Distinct from fully resolved tasks, which require passing every hidden test and remain the ProgramBench primary observation.

Top models (higher is better)

ModelScore
GPT-6 Astra85.4
GPT-6.1 Sol85.3
Gemini 4 Argon83.7
GPT-6 Sol81.9
Claude Haiku 5.578.7
Claude Sonnet 5.576.3
Grok 4.775.7
Grok 4.674.0
GPT-5.6 Terra72.3
Sonnet 572.1
DeepSeek V4.1 Flash72.0
Gemini 3.8 Flash71.9
Muse Spark 1.371.7
MiMo-V2.6-Flash71.7
DeepSeek V4 Pro 081370.1
MiMo-V2.6-Pro69.8
Hy4 Preview69.2
Gemini 3.7 Flash68.7
GPT-5.468.6
GPT-5.6 Luna68.3
GLM-5.366.3
GPT-6 Luna66.2
Grok 4.565.2
Opus 4.764.2
Kimi K362.8
GLM-5.262.6
Muse Spark 1.261.9
Gemini 3.6 Flash59.9
Sonnet 4.659.1
DeepSeek-V4-Flash-073157.3
Gemini 3.5 Flash56.4
GPT-5.4 Mini54.4
GLM-5.150.9
Kimi K2.7 Code49.4
GLM-5.3 Flash48.6
Kimi K2.648.0
DeepSeek-V4-Pro47.8
Inkling-Small42.5
Qwen3.8-Max40.0
Gemini 3.1 Pro Preview39.5
Loading Atlas data…