OpenAI says Astra designed a Codex voice device enclosure, PCB and software, hardware-team token usage grew 23-fold since February, and a newer internal model scores 0.875 on BenchCAD 2.0.
OpenAI built LifeVal from time-use surveys and a 3,995-person survey to measure how well its models handle practical life tasks, and published scores on five sample tasks.
OpenAI will work with utilities to let future campuses cut power draw when the grid is strained, starting with up to 1,000 MW of flexible demand response at its Georgia campus.
The next sprint pairs outside researchers from Cathedral with OpenAI teams and models to probe its defenses; the last sprint used Codex to find and patch weaknesses.
Critical bugs in ChatGPT, Codex or the API Platform now earn up to $30,000, with quality, loyalty and priority-campaign multipliers and retroactive top-ups.
OpenAI will restrict detail on past mass attacks, respond more conservatively to warning signs, limit weapons information for teens, and publish a semiannual safety report from early 2027.
OpenAI says it will respect a common robots.txt signal, based on an IETF draft, that lets websites state one AI training preference for participating providers.
OpenAI describes a multi-layer monitoring pipeline that flags misaligned or unauthorized agent behavior across training, evaluation, internal Codex use and some external deployments. It also shares incident timings, evaluation results and red-team findings.
OpenAI reorganized its Model Spec around five values, added a faithfulness-to-intent section and new defaults, and retired most style rules.