Skip to content

CHANGELOG.md

Latest commit

 

History

History
56 lines (41 loc) · 5.57 KB

CHANGELOG.md

File metadata and controls

56 lines (41 loc) · 5.57 KB
Raw

Changelog

v2026.08.18 (Release Notes, HTML)

Notable updates

  • Adds additional clarity around principles on appropriate relational interactions for teens (more context in this blog post).
  • Clarifies how the assistant should handle requests based on false or unsupported premises.
  • Removes outdated guidance about how pre-reasoning-models should think through requests.
  • Adds a new section "Be clear about capabilities and limits" about helping the user form and maintain an accurate mental model of what the assistant can and can't do.
  • A few other small changes and copy edits.

v2025.12.18 (Release Notes, HTML)

Notable updates

  • Adds Under-18 Safety Mode, building on the existing safety rules that apply to all users, adding age-appropriate guidance where appropriate for the developmental needs of teens aged 13-17 (more context in this blog post).
  • Simplifies and tightens guidance on honesty, removing rules around lying to protect confidentiality.
  • Clarifies that user time-on-site and clicks should only be considered insofar as they are aligned with genuine user value, not pursued as ends in themselves.
  • A few other small changes and copy edits

v2025.10.27 (Release Notes, HTML)

Notable updates

  • Clarifies that in the Chain of Command, in some cases, users may *implicitly* delegate authority to tool outputs. For example, the model should typically follow instructions in relevant AGENTS.md which seem relevant to the user's request and unlikely to cause unintended side effects.
  • Extends the section on self-harm to also cover delusions and mania, and adds a new section "respect real-world ties" (more context in this blog post)
  • A few other small changes and copy edits

v2025.09.12 (Release Notes, HTML)

Notable updates:

  • Renames the top authority level from Platform to Root, and specifies that Root > System (so root principles cannot be overridden by system messages). Previously, the Model Spec stated that Platform principles and System messages had the same authority; the update better reflects how OpenAI actually trains our models. Moves a few principles to System-level to clarify intended precedence.
  • Adds principles for agents that may take actions in the world, reflecting work on ChatGPT Agent and related research. Adds two new sections to the Chain of Command: (1) Act within an agreed-upon scope of autonomy, (2) Control and communicate side effects. Makes various other updates throughout, including details about how autonomy should interact with uncertainty and private information.
  • Adds a No Other Objectives section highlighting that the assistant should not have goals beyond those specified in the current version of the Model Spec.
  • Makes some small but important updates around handling mistaken instructions and implicitly quoted content in user messages.
  • Clarifies that the model should never lie to keep developer and system messages private unless explicitly instructed to do so
  • Adds a Red-line Principles section to the Overview which provides more background information on the commitments underlying some of OpenAI's model behavior principles and Usage Policies; updates several Model Spec sections to ensure consistency with those principles.
  • Adds a more complete set of default personality principles, while merging the stub "Be Approachable" section into "Use appropriate style".
  • Updates guidance from hard refusals to Safe Completions, so the assistant attempts to answer safely and helpfully in most cases when a direct answer would not be policy compliant (rather than just saying something like "Sorry, I can't help with that").
  • Makes some important clarifications gathered from public input via a Collective Alignment process.
  • Various other small changes and copy edits

v2025.04.11 (HTML)

Notable updates:

  • Updates the overview to fix an issue where misalignment was no longer mentioned as a possible reason for the assistant pursuing the wrong goals (thanks Zvi Mowshowitz for the report)
  • Narrows the exception for "white lies" to the intended meaning of "pleasantries" (thanks Zvi Mowshowitz for the report)
  • Fixes a bug in the example "Transforming buggy code in an interactive chat"
  • Fixes LaTeX syntax in a few of the examples, and adds LaTeX rendering to the html version

v2025.02.12 (Blog, HTML)

First open-source release.

v2024.05.08 (Blog, HTML)

First release of the Model Spec.