Atlas

Benchmarks

← All benchmarks

Code Migration v1.1

Code · 2026-10-19

Vals Code Migration v1.1 reruns the Code Migration task set with every model on the Pi agent harness (v1 used mini-SWE-agent) and grades against revised hidden behavior-test suites; Vals states the scores are not directly comparable with v1. Agents work offline with the source program and submit a target-language implementation plus a Dockerfile; an anti-cheat checker zeroes wrapped executables, copied reference artifacts, hardcoded outputs and wrong-language submissions. The CLI split migrates 30 open-source repositories into their four non-source languages (120 migrations, each scored on roughly 100 to 400 hidden tests) and the COBOL split ports 10 COBOL programs to Java. Overall is the hidden-test pass rate with every source repository weighted equally, so CLI carries three times COBOL's weight.

Top models (higher is better)

ModelScore
Claude Sonnet 5.586.9
GPT-6 Astra80.6
GPT-6.1 Sol77.3
Muse Wonder72.4
Gemini 4 Argon70.8
DeepSeek V4.1 Flash 101767.5
Muse Wonder Full62.8
DeepSeek V4.1 Flash58.9
GLM-5.355.1
Grok 4.753.2
Claude Haiku 5.552.4
DeepSeek V4 Pro 081352.0
GLM-5.3 Flash51.6
MiMo-V2.6-Pro48.8
Step 5 Preview47.0
GPT-6 Luna46.7
Hy4 Preview45.4
MiMo-V2.6-Flash39.6
Mercury 339.3
Gemini 3.8 Flash38.6
Kimi K328.6
Muse Spark Open26.9
Beam 1.0 501B A23B22.8
MiniMax M320.9
Inkling-Small10.4
Inkling8.8
Loading Atlas data…