OpenAI's next big AI model reportedly needs more safety work before it can launch
The Information reports OpenAI expects to release one or two new models this year. The big upgraded one may not be among them.
The Information reports that OpenAI employees expect "one or two new models" by the end of the year. It is unclear whether the big upgraded model is one of them. That model still needs work to make sure it behaves as expected.
The Information says the model uses a trick already found in Astra, OpenAI's current top model. The AI runs each step through the same layers several times before it answers. This helps with math and coding and lowers costs. But it moves some of the AI's "thinking" out of readable text, so safety researchers can't follow it.
Under an executive order from President Trump, OpenAI also committed to letting the US government see new models first. That can happen as much as 30 days ahead of outside partners.
From the sourceRead the original at The Decoder →Nothing changes for people using OpenAI's tools today. The reports name no release date, only that employees expect one or two new models by the end of the year.
The facts
- The Information reports that OpenAI employees expect "one or two new models" to ship before the end of the year, and it is unclear whether the major upgraded model will be one of them.
- OpenAI's new flagship still needs alignment work, meaning effort to make its behavior fit what researchers and customers want, and the company must also strengthen the testing setup around it.
- Under an executive order from President Trump, OpenAI committed to giving the US government early access to new models, as long as 30 days ahead of outside partners.
- The new flagship uses looped transformers, the same technique central to Astra, OpenAI's current top model. The model passes each step through the same layers several times before producing the next word.
- The Information reports that OpenAI put a deliberate limit on the number of loops Astra runs, so that its step-by-step reasoning stays readable to people.
- A deleted Microsoft write-up said GPT-6.1 Sol uses the same base model as GPT-6 Sol, with two passes instead of three.
- A person familiar with the figures says OpenAI shifted about ten percent of its overall computing power to alignment and safety work following the Hugging Face incident.
Sources
- The Decoder: OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
Published
More in New models
AmazonNew modelsAyobi’s pickReflection AI's Beam, a free-to-download coding AI, is now available on Amazon's cloud
Reflection AI released Beam, an AI built to write and fix software over long jobs. Anyone can download and change it, and Amazon now offers it to run on its cloud.
New modelsOfficial releaseAyobi’s pickvia Hugging FaceIBM releases AI models that chat, make quick calls and transcribe speech in one
IBM says its new Granite Switch 4.1 models handle conversation, yes-or-no style judgments and speech-to-text from one online connection, using the memory of one model.
New modelsvia The DecoderBlack Forest Labs releases FLUX 3 Image Klein, an image AI for home computers
The new model makes and edits pictures up to 4K size. Black Forest Labs says it runs on an NVIDIA graphics card with 8 GB of memory.