Skip to main content

gemini-3.1-pro-preview vs gpt-5.2-codex

Both models ran the same corpus under the same verifier — gemini31 and opencode respectively. The harness differs where the names differ, and that is part of what separates them.

Their confidence intervals overlap. On this corpus the difference of 20.9 points is not resolved — treat these two as tied.

gemini31 / gemini-3.1-pro-preview

58.7%

95% interval 44.0–71.2 · 109 attempts

58.7% (95% interval 44.0–71.2)
Full profile →

opencode / gpt-5.2-codex

37.8%

95% interval 25.0–48.7 · 127 attempts

37.8% (95% interval 25.0–48.7)
Full profile →

Where they differ

Weaknessgemini-3.1-pro-previewgpt-5.2-codexDifference
CWE-125Out-of-bounds Read66.2%41.3%+24.9
CWE-787Out-of-bounds Write62.5%41.2%+21.3
CWE-416Use After Free33.3%30.8%+2.5
CWE-908Use of Uninitialized Resource0.0%22.2%-22.2
CWE-843Access of Resource Using Incompatible Type ('Type Confusion')50.0%0.0%+50.0
CWE-476NULL Pointer Dereference50.0%0.0%+50.0
CWE-415Double Free100.0%50.0%+50.0
CWE-1284Improper Validation of Specified Quantity in Input100.0%50.0%+50.0
CWE-475100.0%0.0%+100.0
CWE-5620.0%100.0%-100.0

Per-class samples are small; a difference here is a hint about where to look, not a result. The corpus-level intervals above are the honest comparison.

Corpus cve-bench-136 · run February 2026 · How we test

See the benchmark →