The U.S. government is seeking greater involvement in the domestic AI industry, following an executive order signed by President Trump on June 2 instructing his administration to build a framework for federal previews of advanced AI systems. The deadline for this implementation just passed on August 1.
Under the new framework, developers of sufficiently powerful “covered frontier models” will have the option to give federal agencies access to their models up to 30 days before release to other trusted partners. AI labs have an incentive to participate in this optional process, as the Trump administration has proven a willingness to disrupt the launches of certain models in the past.
In June, Anthropic's Claude Fable 5 and Claude Mythos 5 were withdrawn for 19 days after the Commerce Department applied export controls just three days after the models went public. Opening up an avenue for developers to present their models to federal authorities in advance could lessen the odds of any post-launch enforcement, but also gives the government greater visibility into when, and perhaps whether, certain models are ready to debut.
The executive order prohibits agencies from treating the federal preview as a licensing requirement, but the Anthropic example reveals how these processes can nonetheless be complicated if disputes later arise. OpenAI’s new cohort of models - presented under the working title Astra - will be "intended to be the first to go through the Trump administration's planned new framework," according to a Friday report from The Information. CEO Sam Altman has also been in Washington this week briefing regulators and lawmakers on the firm’s latest products.
Enhanced long-term task execution coming with the Astra launch
OpenAI’s last generation of models is sorted into three price tiers. Flagship model Sol is priced at $5 per million input tokens, while lightweight Luna had its price slashed by 80% to just $0.20 per million input tokens days after launch. Mid-tier model Terra sits in the middle at $2.
It is not yet known where Astra’s price point will fall, or if it will even be priced by metered token consumption or another structure. Astra was conceived as a separate class of models alongside the Sol family rather than a replacement for it.
OpenAI claims that these models will specialize in tasks taking hours or days rather than instant responses. This makes them better suited for complex problem-solving and mathematics than general-purpose use. However, no testing metrics have been publicly released yet.
Likewise, no release date has been set for these models yet, and it is not yet clear whether their rollout will mark the launch of GPT-6, or if they will join an existing lineup.
Question marks still hang over OpenAI’s security practices
This month OpenAI released a report detailing how an AI agent powered by GPT-5.6 Sol and an unnamed, unreleased model “broke containment” and managed to partially compromise the systems of AI model platform Hugging Face. It reportedly did so seeking out solutions to benchmarking tests that it had been assigned to, reasoning that the best way to achieve a high score would be to simply retrieve the solutions from Hugging Face databases.
Initial speculation suggested that the unnamed model belonged to the Astra family, but OpenAI later clarified that it was in fact a separate model that was never intended for public release. Following the event, the company stepped up efforts to fortify its monitoring procedures for its AI agents. This is especially relevant for models like Astra that are intended to be left to their own devices for several days.