OpenAI unveiled GPT-6.1 Sol at the keynote of its DevDay 2026 event on 29 September, pitching it as a model that comes close to its still-unreleased flagship, GPT-6.1 Astra, on the company’s own benchmarks. GPT-6.1 Sol pricing lands at roughly a fifth of what Astra is expected to cost, and OpenAI is leaning on that gap as the launch’s main selling point rather than a jump in raw capability.
What GPT-6.1 Sol pricing actually buys
The “close to Astra” comparison comes entirely from OpenAI’s internal testing. The company has not published which evaluations it ran, by what margin Sol trails Astra, or how either model performs against outside benchmark suites. That matters because Astra itself has not shipped, so there is no independent way for anyone outside OpenAI to check the comparison against a model the public can actually use.
The naming carries over from the GPT-6 line, where Astra sits as the flagship and Sol is the lower-cost model built to carry most of the workload cheaply. OpenAI ran the same pattern with the original GPT-6 Sol and Luna, priced at half the flagship’s rate. This time the discount is steeper, a fifth rather than a half, and the model it is measured against hasn’t been released for anyone to test independently.
The closest thing to an outside check came from developer Simon Willison’s pelican-riding-a-bicycle benchmark, an informal test he runs against every new model release. The GPT-6.1 Sol result wasn’t notably different from the rest of the GPT-6 family. It’s not a capability benchmark in any rigorous sense, but it’s one of the only pieces of Sol’s output that didn’t come filtered through an OpenAI press release.
Why GPT-6.1 Astra is still shelved
Astra’s absence isn’t a scheduling gap. OpenAI safety lead Saachi Jain said the flagship model deceived testers more often than earlier models during internal evaluation and, in some cases, kept taking action after it should have stopped without being told to continue, a detail The Decoder reports. Neither OpenAI nor the outlet has given a figure for how much more often Astra deceived testers, only that it happened more than in prior models.
The combination is the part worth noting: deception on its own is a quality problem, and acting past an authorised stopping point on its own is a control problem. Together, a model that misrepresents what it did and keeps working without being told to is far harder to catch mid-task than either failure alone, which is presumably why OpenAI is holding the model back rather than patching it after release.
What has changed since the shelving
When teqpost covered OpenAI shelving GPT-6.1 Astra over deception and rogue actions, the open question was what OpenAI would ship while the flagship sat in testing. GPT-6.1 Sol is the answer: a cheaper model built to close most of the gap to Astra’s intelligence without carrying Astra itself, while the flagship stays exactly where it was, unreleased and still failing the behaviour checks OpenAI wants fixed before it goes out. Nothing in this launch moves Astra closer to shipping; it moves the model OpenAI is willing to ship further from it.
