This briefing covers five verifiable AI developments from 8 p.m. on October 31 to 8 p.m. on November 1 (Beijing time): three rule changes from OpenAI, a loophole in Anthropic's wider watermark rollout, and the next step in xAI's gas-turbine lawsuit.
ChatGPT and Codex share one allowance, so coding eats into chat
In a help-center article updated on October 31, OpenAI said that from December, the usage allowance on Plus, Pro and Business plans will be shared across ChatGPT and Codex, measured in tokens, with stronger models and harder tasks consuming it faster; Instant chat does not count, and remaining allowance can be checked in Settings. Once the allowance runs out, users can wait for a reset, upgrade or buy credits, while the switch date and per-plan allowances remain unpublished, so heavy coders need to rethink their usage.
Claude's watermark reaches four more models, but one setting change defeats it
In an email to API customers on October 31, Anthropic widened coverage of Claude's text watermark, with Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 all set to carry it. On Opus 4.6 and Sonnet 4.6, the watermark only applies when the three sampling parameters — temperature, top_p and top_k — are all left at their defaults, and changing any one of them defeats it. Anthropic advised customers who rely on the watermark to keep the defaults, and its detection API remains in limited preview.
xAI gas-turbine case moves forward as judge denies dismissal
A federal judge denied xAI's motion to dismiss on October 31, allowing the NAACP's lawsuit against xAI to proceed. The complaint says xAI operated at least 35 gas turbines at its Colossus data center outside Memphis without the required air permits, violating the Clean Air Act. The judge held that the plaintiffs have standing and that the turbines are not 'mobile sources' exempt from permits. The ruling does not decide ultimate liability, and xAI must keep defending the permitting dispute.
Daybreak Blue session data will be used for training; opt-out still pending
Reporting on October 31 said an email from OpenAI to Trusted Access members stated that from November 29, Daybreak Blue session data will be used to train models even when users have switched training off; such data had previously been excluded entirely. After pushback, OpenAI promised a separate opt-out, with the mechanics still unexplained. The rule covers only Daybreak sessions on personal accounts, and content flagged in safety reviews can still be used for safety purposes.
OpenAI's Evals platform goes read-only and shuts down November 30
Under OpenAI's published schedule, existing evals on the Evals platform became read-only on October 31, with the dashboard and API set to shut down on November 30; the platform was announced for deprecation on June 3. Read-only means teams can only view past evals and must export and migrate before the shutdown. OpenAI recommends Promptfoo as the migration path, and Agent Builder and reusable prompt objects are set to close the same day.
