OpenAI is developing a model family called Astra that it says can coordinate multiple AI agents on difficult problems for hours or days. The company has now used an internal version of the system to produce solutions to ten previously unresolved problems in mathematics and theoretical computer science, according to a report from The Decoder.

The work offers the clearest public indication yet of OpenAI’s attempt to move beyond short, interactive answers toward long-running research workflows. Astra has reportedly been demonstrated by CEO Sam Altman to policymakers in Washington, while the models remain under testing and have no announced public release date.

Astra’s reported focus: extended reasoning

The Astra project is intended to combine several agents that can divide, investigate, critique, and revise work on a complex problem. That design would place it in a different category from conventional chatbots and coding assistants, which generally complete tasks within a shorter interaction window.

The Decoder reported that Astra could eventually appear as GPT-6 or as a future GPT-5 variant, but OpenAI has not decided how it will name or package the system. The project is also described as a new model class alongside OpenAI’s reported Sol, Terra, and Luna families.

The company’s longer-term objective is to build systems that can plan, reason, experiment, and continue working without constant human intervention. OpenAI chief scientist Jakub Pachocki has previously discussed the goal of AI systems capable of working on problems for hours or days, according to The Decoder. Astra appears to be an effort to turn that ambition into a tested product direction.

Ten mathematical results, with human verification

OpenAI’s math report says an internal Astra system solved ten open problems in areas including high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. The problems had reportedly resisted progress for at least a decade, and in many cases considerably longer.

One reported result establishes the existence of non-sofic groups, resolving a major question in group theory. The mathematical significance of the results is being assessed by researchers rather than determined solely by OpenAI. Thomas Bloom, a University of Manchester mathematician who runs erdosproblems.com, described the constructions as major news in comments reported by The Decoder.

OpenAI says the model generated the mathematical arguments, while researchers helped convert them into papers and formalize every proof in Lean. That formalization produced machine-checkable certificates, offering an important layer of verification beyond a plausible-looking natural-language answer.

The distinction matters. OpenAI said its researchers take responsibility for the accuracy of the published work, but that the core arguments came from Astra. The company has also argued that assigning full human authorship to a proof generated by an AI system would misrepresent the system’s contribution. The final status of the results will depend on scrutiny from mathematicians and other independent experts, not only on the model’s ability to produce Lean artifacts or explanatory walkthroughs.

What the evidence does and does not show

The strongest performance claims in this story come from OpenAI’s own report, as relayed by The Decoder. There is no independent benchmark in the available evidence showing that Astra consistently outperforms other advanced systems on long-horizon research tasks.

OpenAI estimated that the tokens used to generate all ten solutions would have cost about $2,000 at Sol’s API rates. That figure is a vendor estimate and does not include the wider cost of developing, operating, supervising, and validating the system. It also does not establish that the same approach would be economical across a large number of research problems.

Noam Brown, a researcher associated with the test-time reasoning technology used by Astra, said the system had not solved any of the seven Millennium Prize Problems. He also suggested that substantially more test-time compute could improve results. His comments point to both the promise and the limitation of the approach: difficult reasoning may be possible, but only with significant inference time and resources.

The results therefore should not be read as evidence that AI has replaced mathematicians. Bloom rejected that interpretation, noting that the system relies on more than a century of mathematical knowledge and on work by the researchers who built and trained it. Human experts remain involved in problem selection, evaluation, formalization, and publication.

Why Astra matters to AI builders and enterprises

For AI builders, Astra highlights a shift from model quality measured by single responses to systems judged by their ability to sustain a workflow. A useful long-running agent must preserve goals, manage growing context, detect failed approaches, and decide when to ask for human intervention. Those requirements are much closer to operating a research team than to generating text on demand.

The multi-agent design also creates new reliability risks. The Decoder noted that errors can compound as a process continues and that coordination overhead can make multiple agents less effective on tightly connected planning tasks. A system that produces a convincing but incorrect conclusion after several hours could be more difficult to audit than one that fails quickly.

That makes the Lean formalizations particularly relevant for high-stakes technical work. Machine-checkable proofs do not solve every question about research quality, but they provide a concrete validation mechanism. In other domains, builders will need comparable controls: reproducible execution, source tracking, intermediate checkpoints, tests, and clear escalation paths for uncertain results.

Enterprise buyers should also watch the cost profile. A system that spends hours or days on a task may create value in scientific discovery, engineering, or complex analysis, but its economics will differ sharply from ordinary chat or retrieval workloads. OpenAI’s reported token estimate is useful context, not a complete operating-cost model.

Government review and deployment uncertainty

The Decoder reported that Astra models are expected to be among the first systems considered under a planned U.S. government review framework requiring approval before public release. The timing and final structure of that framework remain uncertain in the available evidence.

A review process could affect how OpenAI deploys long-running AI agents, particularly if they can conduct research, write software, or operate with limited supervision. It may also establish a precedent for treating extended autonomous workflows differently from ordinary model releases.

OpenAI has not announced whether Astra will be offered through an API, embedded in ChatGPT, released as a research preview, or kept internal. Until those decisions are made, developers cannot assess its latency, pricing, tool access, safeguards, or practical limits.

What to watch next

The immediate signal will be OpenAI’s full math report and whether independent mathematicians validate the ten results. The quality of the Lean formalizations, the amount of human intervention, and the reproducibility of the reported solutions will be more informative than the headline count alone.

The market should also watch for a public Astra evaluation covering long-running tasks rather than isolated demonstrations. Useful measures would include success rates, recovery from failed reasoning paths, compute consumed, and performance against single-agent alternatives.

Finally, OpenAI’s product naming, release channel, and government-review status will clarify whether Astra is an experimental research system or the foundation of the company’s next mainstream model generation.

Creati.ai perspective

Astra’s importance is less about ten mathematical answers than about the operating model OpenAI is pursuing: AI systems that spend substantial compute and time on a problem, coordinate specialized processes, and return work that can be checked. That could expand the role of AI in research, but only if reliability and cost improve alongside raw reasoning capability.

For builders, the practical lesson is to treat long-running agents as systems engineering projects. Persistent state, verification, resource controls, and human review will matter as much as the underlying model. OpenAI’s claims are significant, but Astra’s real test will come when independent users can measure whether it completes difficult workflows consistently outside a carefully selected demonstration.