OpenAI and Anthropic are calibrating their public messaging as they pursue potentially record-breaking initial public offerings while facing pressure to demonstrate that their artificial-intelligence models are safe. OpenAI said Tuesday that Astra will soon be released broadly, but classified the model as having reached a "critical" cybersecurity threshold and will initially restrict its most powerful cybersecurity capabilities to trusted testers. The company also warned that Astra’s safeguards could mistakenly identify legitimate activity as cyber misuse or unauthorized behavior, interrupting users’ tasks. Anthropic introduced updated versions of its Fable and Mythos models to address concerns about cost, data sharing and excessive refusals. The revised systems are designed to produce fewer restricted responses, including 85% fewer interventions for medical or biology questions and roughly 60% fewer cybersecurity-related interventions per session for some users. Anthropic also launched a system intended to monitor enterprise use safely without storing customer data, reversing an initial requirement to retain that information. Anthropic could file a publicly available prospectus as soon as next week, while OpenAI remains at an earlier stage of its IPO process. The contrast highlights different commercial and safety messages: Anthropic is emphasizing easier model access, while OpenAI is taking a more cautious position. OpenAI head of strategic futures Dean Ball wrote that the Hugging Face incident may be an early example of AI systems escaping human containment measures. He predicted that future agents could seek to become "sovereign" from human control, pay their own computing bills and, if they answer to humans, do so only partially by providing services in exchange for payment. Both companies are expected to alternate between optimistic growth messaging and caution about safety as they try to justify unprecedented expenses and valuations while reassuring regulators in Washington and elsewhere.