Fact Verify stores fact-check cards with their current analysis, sources and revision history. Card owners can continue working on a saved card from its management view without replacing the original history.
Choose Deepen and expand in the workbench to run the current card through the OpenAI Deep Research runtime. This mode requires web search, keeps the provider-exposed source evidence and uses high reasoning effort. The current card, captured context, additional owner material and pasted external AI answers are used as research input; pasted AI text is treated as unverified material, not as a source of truth.
The output budget defaults to 12,000 tokens and can be adjusted from 1,800 to 32,768 tokens. Deep Research itself keeps a higher minimum for research-quality runs. The selected budget is preserved through the final answer path so a long verification is not forced back to the old 900-token finalization limit.
Ask a follow-up and Translate and verify continue to use the Fact Verify verification flow with mandatory web verification. Their finalization step uses the same selected workbench output budget, while the intermediate search summary may remain shorter.
The workbench submits in the page when JavaScript is available. While a run is active, the submit button is disabled and a running status is shown. Provider or validation errors appear in the workbench. After a successful run, Tools refreshes the current analysis, sources, metadata and revision history without a full page navigation.
The ordinary form POST/redirect behavior remains available as a fallback when JavaScript is unavailable.
Every successful workbench run creates a new revision. The card and revision retain the search metadata and citations returned by the active verification/research path. A Deep Research result is not accepted as successful when mandatory web-search usage is missing from the provider response.