Sharing the latest Model Spec

• OpenAI releases major Model Spec update defining required model behavior.
  • Builds on May 2024 original; incorporates external feedback + alignment research.
• Six core principles govern all models: command chain, truth, quality, bounds, tone, style.
  • Priority order: OpenAI platform → developer → user.
• Intellectual freedom explicit: controversial topics allowed, arbitrary restrictions banned.
  • Hard limits remain: terrorism facilitation, privacy violations, serious harm.
• Spec released CC0 — full public domain, freely adaptable.
  • Eval prompts + source code on GitHub for community use.
• Benchmark shows significant adherence gains vs. May 2024 systems.
  • Test set: model-generated + expert-reviewed prompts, typical + edge cases.
• Community pilot: ~1,000 individuals reviewed rules and model behavior.
  • OpenAI acknowledges narrow sample; broader input mechanisms planned.
• Versioned spec now lives at model-spec.openai.com; no separate release posts.

OpenAI (no individual authors; published February 12, 2025) · ** · Read on openai.com