OpenAI could unveil a new generation of artificial intelligence models as early as Friday, November 6, 2026, according to a report by The Information, with updated Astra and Sol models expected to headline the launch. The report arrives days after the company published benchmark results showing an unreleased internal model outperforming its current flagship on demanding engineering tasks.
The launch would continue a rapid run of releases where reasoning depth has become the AI race’s central battleground. Yet the most important details remain unconfirmed. OpenAI has not announced the new models by name, has not set out their pricing, and has not said when or how they will be made available. The Friday date rests on reporting and claims on social media, rather than on any official announcement.
New Internal Model Could Be Much More Powerful
The clearest evidence sits in the benchmark figures OpenAI published on November 3. In those results, an unreleased internal model achieved an average score of 87.5 percent on BenchCAD 2.0, a test that measures how well an AI system handles engineering work involving three-dimensional models, assemblies and circuit boards. GPT-6 Astra, the current leading model, managed 59.4 percent even at its maximum reasoning setting.
That gap matters because engineering benchmarks test far more than general knowledge. Designing a workable three-dimensional part, checking how components fit together in an assembly, or laying out a circuit board all require a system to hold many constraints in mind at once. A model that scores well here can handle the structured, multi-step problems engineers face daily, where one error can ruin a design.
A breakdown shared on X by user @deredleritt3r added detail. According to that post, the internal model scores around 76 percent at its lowest estimated setting, above the 59 percent Astra reaches at its best. Astra climbs from 14 percent at 44 cents per task to 31 percent at 2.73 dollars, 41 percent at 5.11 dollars, 45 percent at 6.67 dollars and finally 59 percent at 11.05 dollars. The internal model, by contrast, is estimated to reach around 76 percent at roughly 11 dollars per task, around 83 percent at 23 dollars, around 86 percent at 34 dollars and around 88 percent at 37 dollars.
Those numbers come with important caveats. The user who shared them said the costs were extrapolated from a graph in OpenAI’s own benchmark publication, and that the labels for the reasoning settings were estimates rather than official figures. They describe the estimated cost of completing benchmark tasks, not confirmed prices for using the model through OpenAI’s programming interface. Nor do they prove the internal model beats Astra or Sol on every kind of task. A strong engineering result is evidence of real progress in one area, not proof of superiority across writing, coding, science or everyday conversation.
New Sol Model Could Receive Major Training Upgrades
The same post on X also claims OpenAI is preparing another Sol model that will benefit from the training methods used to build its more capable internal system. According to the claim, the coming Sol model will not simply be a distilled version of the larger system, in which a smaller model is trained to imitate a bigger one. Instead, it will reportedly go through similar training itself, using games and intensive simulated environments designed to sharpen reasoning and problem-solving. If accurate, the new Sol model could improve substantially on GPT-6.1 Sol, particularly on tasks needing sustained reasoning. OpenAI has not confirmed the training details or published any benchmark comparison with its predecessor, so the upgrade remains an unconfirmed claim for now.
OpenAI Reportedly Planning Friday Release
The Information’s report of November 4 says OpenAI could introduce new Astra, Sol and cybersecurity-focused AI models as early as November 6. The X post goes further, claiming the internal model from OpenAI’s engineering research will also be available on Friday. The Information has not confirmed that the benchmarked internal model is among Friday’s releases, and OpenAI has said nothing about names, pricing or availability.
The speculation comes against a backdrop of quickening change. OpenAI launched GPT-5 on August 7, 2025, as a unified system with a fast model for everyday questions, a deeper thinking model for harder ones and a real-time router choosing between them. Earlier this year, in February 2026, the company retired several older models from ChatGPT, underlining how quickly each generation now gives way to the next.
If Friday does bring new models, the questions that matter most will be practical ones. Will the internal model that scored 87.5 percent on BenchCAD 2.0 actually be among them. How much will the new systems cost to use, and will the steep price of high reasoning settings put the best performance out of reach for ordinary users and small developers. And how will they perform beyond the engineering benchmark, on the coding, science and everyday tasks where most people will meet them. The published results suggest OpenAI has made a genuine leap in at least one demanding area. Whether that leap reaches users on Friday is something only the company can now confirm.
