Artificial Intelligence
OpenAI to ship models as flagship remains unfinished
The Information, citing employees and other people familiar with the work, said OpenAI expected to release one or two models by year-end, while its main upgraded model still needed alignment and testing work.
- The Information, citing employees and people familiar with the projects, said OpenAI would likely release one or two new models before the end of the year.
- The article said the company’s major upgraded model still needed alignment work and better testing infrastructure.
- It said the model used looped transformers, the same technique as Astra, and that the method improved maths and coding performance but made safety monitoring harder.
- Under President Trump’s executive order on advanced AI evaluations, OpenAI agreed to give the US government access to new models up to 30 days before outside partners and longer before public launch.
- Microsoft later removed a technical write-up about GPT-6.1 Sol after it described looping inference and the base model used by GPT-6 Sol.
The Information, citing employees and people familiar with the projects, said OpenAI was expected to release one or two models by year-end. It was unclear whether the company’s main upgraded model would be among them, and the article said that model still needed alignment work and stronger testing infrastructure.
The article said OpenAI’s next flagship used looped transformers, the same method as Astra, its current leading model. It said the design improved maths and coding performance and lowered costs, but also pushed some reasoning into internal representations that people reviewing the model could not inspect.
OpenAI had shifted about a tenth of its compute towards alignment and safety after the Hugging Face incident, according to a person familiar with the figures. The Information said that equated to hundreds of megawatts of data-centre capacity, although the current split was not clear.
Under President Trump’s executive order on advanced AI evaluations, OpenAI agreed to let the US government inspect new models up to 30 days before outside partners and longer before public launch. Microsoft later took down a public technical page on GPT-6.1 Sol, a cheaper Astra alternative, after it described looping inference and the base model used by GPT-6 Sol.
OpenAI chief scientist Jakub Pachocki said chain-of-thought monitoring had become fragile and was moving in a negative direction. He also warned against an industry-wide push towards models without readable reasoning.
Named in this story
People
- President Trump
- issued the executive order governing advanced AI evaluations
- Jakub Pachocki
- said chain-of-thought monitoring was becoming fragile
- Noam Brown
- said about ten per cent of his team also worked on security and safety
Companies
- OpenAI
- was developing the new flagship model and planning releases
- Microsoft
- published and then removed a technical write-up that disclosed looping inference details
Governments and agencies
- United States government
- was due to receive advance access to new models
Products and systems
- Astra
- was OpenAI’s current top model and used looped transformers
- GPT-6.1 Sol
- was the cheaper Astra alternative described in Microsoft’s write-up
- GPT-6 Sol
- was the base model Microsoft said GPT-6.1 Sol used
How the source tells it
The piece kept a reportorial register, but it leaned on urgency, unnamed sourcing and safety concern to frame OpenAI’s schedule and model design.
- urgency repeated time pressure around year-end release and racing to finish the model
- uncertainty likelihood language and sourcing from employees and people familiar with the work
- safety concern alignment, testing and monitoring difficulties were presented as blockers