Parent: #64227

Summary

Add the shared benchmark harness and release gate for GPT-5.4 vs Opus 4.6 so parity claims are evidence-backed.

Scope

• shared task suite
• completion / unintended-stop / valid-tool-call metrics
• targeted auth/proxy mislabel regressions
• release gate document and CI entrypoint if feasible

Acceptance

• same prompts, tool surface, sandbox policy, and continuation policy for both models
• GPT-5.4 must match or beat Opus 4.6 on completion rate
• equal or lower unintended-stop rate
• equal or higher valid-tool-call rate
• zero fake-success cases
• zero mislabeled auth/proxy/DNS failures