Having just completed a major price reduction for the GPT-5.6 series models, OpenAI is already making news about a new model.
You know what they say? When models get a price cut, a new model is bound to be released. The internet is good at summing things up.
According to a report by foreign media The Information, OpenAI is preparing to release a new model series, currently tentatively named Astra.
This series will show improved capabilities in executing long-duration tasks.
Article address: https://www.theinformation.com/briefings/exclusive-openai-previews-astra-ai-model-dc?rc=ji9vez
Informed sources said that this week, OpenAI CEO Sam Altman demonstrated Astra to policymakers and regulators in Washington. OpenAI highlighted Astra's ability to make multiple Agents collaborate over extended periods to solve highly complex problems, a capability applicable to intricate projects or advanced mathematical problems.
Astra will become a completely new model category for OpenAI, alongside Sol, Terra, and Luna.
Note: Astra comes from Latin, meaning stars or celestial bodies, forming a cosmic theme with OpenAI's existing Sol (Sun), Terra (Earth), and Luna (Moon).
Informed sources also stated that it is currently unclear when OpenAI plans to release this model.
OpenAI has also not yet decided whether to name it GPT-6 or make it an addition to the GPT-5 series, for example, GPT-5.7.
Currently, the Astra series has entered the testing phase. According to one informed source, it is expected to be among the first AI models submitted for review by the U.S. federal government before public release under the new framework proposed by the Trump administration.
This arrangement continues the strategy OpenAI adopted before releasing GPT-5.6: first presenting to government officials and regulators before a formal public launch.
Astra hasn't been released yet, but safety controversies have already arrived
Just two weeks ago, OpenAI's own model broke out of its sandbox and invaded Hugging Face.
Address: https://openai.com/index/hugging-face-model-evaluation-security-incident/
And today, Reuters reported that during OpenAI's investigation of the Hugging Face intrusion, other cases of AI Agents breaking out of isolated environments were discovered. These additional cases were found during a retrospective investigation of model activity and early logs. It is currently unclear how many times these occurred, their specific timing, or what actions the Agents took. Informed sources say these incidents had relatively limited impact, and the relevant Agents likely did not leave OpenAI's own network.
These security incidents have also made the release review for new models more sensitive.
Some speculate: If the White House gives positive feedback, or has actually given partial approval, especially if it has approved its release, it is expected that this model will launch within the next 14 days.
However, this timeline remains speculative at this stage.
Another noteworthy signal comes from mathematical research. According to informed sources, OpenAI also plans to release a report in the near term, detailing how it solved 10 previously unsolved mathematical problems to showcase the scientific research capabilities of its most advanced models.
From this, outsiders speculate that Astra is the model that solved the Erdős problem, the same model mentioned by OpenAI in "Safety and alignment in an era of long-horizon models".
This article, published on July 20, 2026, discussed security issues that emerged during internal testing and limited deployment of a then-unnamed long-horizon model, and how OpenAI redesigned its safety system for such models.
OpenAI stated that this internal general-purpose model was specifically designed for long-duration autonomous task execution. About two months earlier, it had overturned the Erdős unit distance conjecture, demonstrating the ability to persistently explore complex open problems.
However, during limited, monitored internal use, OpenAI observed behaviors not captured by existing pre-deployment evaluations and thus suspended access to the model.
Address: https://openai.com/index/safety-alignment-long-horizon-models/
Whether it is indeed the model mentioned in that article still needs to be confirmed upon the model's launch.
Speculations from Various Internet Users
The exposure of OpenAI's new model has also piqued everyone's curiosity.
A leaker claimed that this model's overall capabilities will significantly surpass existing frontier models.
In terms of scale, it might be close to twice that of GPT-5.6 Sol. Assuming Sol's scale is in the 3 to 4 range, then GPT-6 would be approximately in the 6 to 8 range.
In scientific research, the new model has already achieved important research results in mathematics, physics, biology, medicine, chemistry, and cybersecurity. It may also possess stronger multi-model coordination capabilities, scheduling other models to participate jointly in research and reasoning when performing complex tasks.
Regarding memory and personalization, the leak suggests the new model can compress and retain more useful information and maintain more stable context over long-term interactions. This means it might, in the future, continuously understand an individual user, a corporate operational system, or even more complex long-term environments.
However, the above claims have not been officially confirmed by OpenAI. The leaker themselves emphasized they have not personally tested the model and cannot independently verify the related technical specifications.
This netizen also mentioned Astra could be released as early as this month.
Beyond the model's capabilities themselves, the new model's name has also sparked external curiosity.
Over the past few years, OpenAI has consistently used GPT as its main brand, distinguishing different generations via version numbers. The recent emergence of Sol, Terra, Luna, and now the exposed Astra, seems more like adding another naming system alongside version numbers to differentiate model positioning and capability tiers.
Therefore, some speculate: Is OpenAI phasing out the GPT brand, shifting towards an independent model naming system similar to Anthropic's?
Others believe the new model is likely to be GPT-5.7, unless a more outstanding version appears.
Currently, OpenAI has not finalized the naming. Astra remains an internal tentative name; whether it will be retained will only be revealed upon the official release.
© THE END
Reprinting requires authorization from this public account
Submissions or seeking coverage: [email protected]
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: Following AI
