• OpenAI set out a playbook on May 29, 2026 to steer independent third-party evaluations of frontier model capabilities, safeguards, and model-to-model comparisons.
• Evaluation reports will be pushed to state the claim being tested, then document the harness, tools, safeguards, and budget used to elicit results.
• OpenAI will ask capability evaluators to use Codex as a baseline agent interface to reduce under-elicited performance in tool-using, multi-step tasks.
• Plans include broader sharing of maximum-elicitation guidance, plus access to reasoning traces to assess deception, sandbagging, or evaluation awareness.
• OpenAI will prioritize research on how harness design, context management, tool access, retries, and resource budgets shift measured capability or safeguard robustness.

Disclaimer: This news brief was created by Public Technologies (PUBT) using generative artificial intelligence. While PUBT strives to provide accurate and timely information, this AI-generated content is for informational purposes only and should not be interpreted as financial, investment, or legal advice. OpenAI Inc. published the original content used to generate this news brief on May 29, 2026, and is solely responsible for the information contained therein.