OpenAI’s next large model is again the subject of fresh speculation. According to X user Leo, cited by ABMedia, OpenAI is preparing to release a new model family codenamed Astra as soon as next week.
The leak describes Astra not as a small update to an existing model, but as a new pretraining foundation model. Its internal development codename is said to be “mewfour,” and it has reportedly entered testing. OpenAI has not officially confirmed the information.
After the claim surfaced, several well-known AI community accounts, including Mark Kretschmann, reposted it.
Astra said to differ from recent GPT and o-series updates
ABMedia said Astra would be different from OpenAI’s GPT-5, GPT-5.5 and o-series reasoning models released over the past year. Models such as GPT-5.5, o3 and o4 were described as systems that improved on existing foundations through post-training, reinforcement learning, tool use and reasoning optimization.
A new pretraining model would mean OpenAI trained a system again from the ground up with larger-scale data, compute and architecture. The report said that kind of shift usually points to a step change in knowledge capacity, generalization, multimodal understanding and reasoning, rather than a simple feature refresh.
Questions emerge over potential regulatory treatment
The market is also watching whether Astra could face the same kind of regulatory outcome that affected Anthropic Claude Mythos 5.
In June, the U.S. Department of Commerce told Anthropic to suspend access for foreign persons to advanced models including Mythos 5 and Fable 5, citing national security and cyber offense and defense concerns. The report described it as the first frontier AI model case affected by export controls.
The U.S. government lifted those restrictions in July, and Anthropic resumed global service.
OpenAI research paper already named Astra
OpenAI recently published research titled “Ten advances in mathematics and theoretical computer science,” and the paper included Astra by name for the first time.
The paper summarized 10 major advances by OpenAI in mathematics and theoretical computer science, covering algebra, combinatorics, geometry, probability, number theory, analysis and theoretical computer science.
Among the results mentioned were solving an open mathematics problem that had remained unresolved for decades at a cost of $2,000, producing new proofs in graph theory and discrete mathematics, and improving the theoretical upper bound for matrix multiplication algorithms.
The research also cited new results in quantum information, coding theory and computational complexity.
More importantly, OpenAI said in the paper that these research outcomes were produced by “our next major model Astra.” ABMedia said that amounts to the company’s first formal public confirmation that Astra is its next core model family, and that the model is already capable of contributing to frontier mathematics research.
OpenAI said the research had been submitted for academic peer review and that publishing the results was intended to support mathematics and theoretical computer science research while showing the scientific reasoning ability of the new generation of AI models.
Kalshi market prices point to an August release
Beyond the social media leak, U.S. prediction market Kalshi has listed contracts tied to Astra’s release date. At the time cited in the report, the market implied the following odds:
- Release before Aug. 14: 33%
- Release before Aug. 21: 73%
- Release before Sept. 1: 76%
Those figures show that traders expect OpenAI to formally unveil Astra within the next few weeks.