OpenAI updates Model Spec on teen interactions, unsupported premises and capability transparency

What changed

OpenAI says it updated the Model Spec on August 18, 2026. The release notes highlight clearer guidance on appropriate relational interactions for teens, more explicit handling of false or unsupported premises, removal of outdated guidance written for pre-reasoning models, and a new section titled Be clear about capabilities and limits. OpenAI also notes additional smaller edits across the specification.

Why it matters

These changes are useful for developers and product teams because they make several model-behavior expectations more explicit. Handling unsupported premises well can reduce the chance that an assistant simply accepts a user's mistaken assumption. Clear capability statements can also reduce confusion about what a model can actually access, remember, verify or accomplish. The teen-interaction clarification is especially relevant for products serving younger users, where relational framing and dependency risks require additional care.

What teams should review

Teams building assistants should compare their own system prompts, evaluation sets and safety tests against the current Model Spec rather than relying on older snapshots. In particular, it is worth testing whether an assistant challenges unsupported claims appropriately, avoids implying capabilities it does not possess, and follows age-appropriate relational boundaries. The Model Spec is guidance for intended model behavior; production teams still need product-specific safeguards, monitoring and evaluation around their own use cases.