OpenAI previews Astra model built to coordinate long-running agents
Sam Altman demonstrated the unreleased model to policymakers as a federal review process approaches its August 1st deadline.
By Ryan Merket · Published
Why it matters
Astra would push OpenAI from parallel agent features toward models built around persistent coordination, while making federal review a direct part of the release process.
OpenAI CEO Sam Altman demonstrated an unreleased model family called "Astra" to policymakers and regulators in Washington this week, pitching a system designed to coordinate multiple AI agents over long periods, The Information reported.
Altman's choice of audience places Astra inside the tightening relationship between frontier AI development and Washington. The model is already being tested, according to the report, and OpenAI intends for it to become the first of its models reviewed under a new federal process for examining powerful systems before release.
OpenAI has not settled on the final product name. The Information reported that Astra could become GPT-6 or another entry in the GPT-5 line, such as GPT-5.7. The tentative Astra designation describes a new class of models alongside Sol, Terra and Luna, the three capability tiers OpenAI introduced with GPT-5.6 on July 9th.
From parallel agents to persistent work
Astra's central pitch is coordination. OpenAI told officials that multiple agents could work together on difficult projects or advanced mathematics over an extended period, according to The Information. That moves the competitive target beyond producing a stronger answer to a single prompt. Astra is being developed for work that must be divided, delegated, checked and carried forward across a longer timeline.
OpenAI has already started shipping pieces of that approach. GPT-5.6's "ultra" setting coordinates agents across parallel workstreams, while the model's application programming interface includes a beta feature that runs concurrent subagents and combines their results. Astra would need to improve persistence and coordination substantially to merit a separate model class rather than another GPT-5.6 feature.
The demand case comes largely from OpenAI's own usage data. In a June report on how agents are changing work, OpenAI said more than 70% of sampled Codex users in May requested at least one task estimated to require more than an hour of human work. By June, users at the 99th percentile were generating more than 60 hours of agent turns per day by running multiple agents in parallel. OpenAI described those human-time estimates as directional rather than exact.
Those figures explain the product bet. A model that can reliably supervise several agents for hours or days could take on software migrations, research projects and analytical work that currently require a person to keep restarting, correcting and reconciling separate runs. Reliability becomes harder to measure as the task grows: a system can produce many plausible intermediate steps while still failing the final objective.
Washington gets an early look
Astra is also arriving as the federal government builds a formal path into frontier-model release decisions. President Donald Trump's June 2nd executive order directed agencies to create a voluntary framework under which developers could provide the government with access to a "covered frontier model" for as long as 30 days before releasing it to trusted partners.
The order gave officials 60 days to establish benchmarks for identifying covered models and design the submission process, putting the deadline on August 1st. It explicitly says the framework does not create mandatory licensing or preclearance. The practical pressure on frontier labs could still be significant when federal agencies control procurement, security reviews and access to government partnerships.
The framework is focused on advanced cyber capabilities. The National Security Agency would help determine whether a model crosses the covered-model threshold, while the Commerce Department's Center for AI Standards and Innovation conducts model evaluations and develops testing methods. The Information reported that Astra is intended to pass through the new process, though the model's release schedule remains undecided.
That makes Altman's Washington demonstration part product preview and part regulatory preparation. OpenAI is giving officials an early view of a model whose agent coordination could expand both productive use and the scale at which a system searches for vulnerabilities, writes code or pursues an incorrect plan.
The math claim will draw scrutiny
OpenAI also plans to publish a report describing how its advanced AI solved 10 previously unsolved mathematics problems, according to The Information. The report is expected before Astra's release, and the source story did not identify the problems or the researchers who verified the results.
OpenAI has previously documented models contributing to new proofs, locating overlooked literature and helping mathematicians close open problems. Its November 2025 report on science work with GPT-5 also warned that models can hallucinate citations and proofs, miss technical subtleties and reproduce existing work without identifying the original source. Independent expert verification will determine whether the coming results represent autonomous solutions, human-AI collaboration or rediscovery of prior work.
Astra's release timing is open, but the direction is clear. OpenAI is building past the single-agent model and toward systems that allocate work among multiple agents, maintain state and operate across longer time horizons. The Washington preview shows that the deployment process for those systems is becoming part of the product cycle itself.