OpenAI Rewrites Model Spec Around Five Values

The values are mostly a reframing; the consequential change is moving more control toward developers and users. That flexibility is useful for agent builders, but it also makes scope design and observability less optional.

• –Side-effect minimization becoming a default makes agent behavior easier to customize, but raises the risk of poorly bounded actions.
• –Removing sub-agents from the high-risk list lowers orchestration friction, while delegation can still multiply cost, data exposure, and tool risk.
• –Dropping the Markdown default gives applications more control over output formats and presentation.
• –The new emphasis on context, memory, human agency, and oversight points toward models making more judgment calls around developer intent.
• –OpenAI frames the document as an intended behavioral target, not proof that deployed models consistently meet it.