OpenAI unveils GPT-6.1 Sol: performance close to its Astra model at a fifth of the price

OpenAI announced GPT-6.1 Sol through a short thread on its official X (formerly Twitter) account, published on September 29, 2026. It is worth noting from the outset that the only available source is that thread of short social media posts —not an article, an extensive technical release or a document with benchmark tables— so this summary sticks strictly to what the company wrote there, without filling gaps or assuming figures or comparisons that were not made explicit.

The main message defines GPT-6.1 Sol as a model that offers "intelligence close to Astra" —a reference to GPT-6 Astra, presumably the high-end model in OpenAI's GPT-6 family— but at one-fifth the price. The company itself presents it as "the most cost-efficient model for its level of performance available today"; in other words, the selling point centers on the capability-to-price ratio, not on being the most powerful model in the catalog.

On pricing, OpenAI details in the thread that GPT-6.1 Sol's cached input costs $0.10 per million tokens, a figure the company describes as 95% cheaper than the standard input price and 50% cheaper than GPT-6's (the text is cut off with a "Show more" at exactly that point, so it is not confirmed which specific GPT-6 variant the comparison refers to). The stated goal is to make frontier intelligence more affordable, both for users' everyday use and so that developers can build and run applications at greater scale.

On performance, OpenAI claims GPT-6.1 Sol represents "a significant improvement" over its predecessor, GPT-6 Sol, in coding, computer use and complex professional work, approaching GPT-6 Astra's results on "several benchmarks" but at a substantially lower cost. The thread does not specify which benchmarks those are, nor does it offer numerical scores, so it is not possible to detail the exact magnitude of that improvement beyond the qualitative characterization the company itself provides.

On safety and alignment, OpenAI maintains that GPT-6.1 Sol shows "major improvements" over GPT-6 Sol in its internal alignment evaluations, again coming close to GPT-6 Astra in this respect. According to the announcement, the new model is "more transparent about its own limitations" and "more reliable at respecting user intent and safety constraints". Once again, these are the company's own claims, with no methodology or results published in this thread.

On availability, OpenAI says GPT-6.1 Sol is accessible from the day of the announcement to all users on the Plus, Pro, Business, Enterprise and Edu plans, both in what the thread calls "ChatGPT Work" and in Codex. The post also links to an entry on the openai.com blog titled "Introducing GPT-6.1 Sol" for anyone wanting more information, although the content of that page was not part of the material analyzed here.

The thread drew considerable attention, with close to 2.8 million views according to the figures shown on the main post, although the source text does not unambiguously label what each of the other engagement counters (replies, reposts, likes, bookmarks) corresponds to, so they are not broken down here in order to avoid attributing a meaning that is not confirmed. Among the visible reactions, one user critically flagged OpenAI's release pace with the phrase "Lets slow down", noting that the company had launched seven new models in the space of a few weeks. That figure of "seven models" is a claim by an outside user in the comments, not a fact confirmed by OpenAI, so it is recorded here as evidence of the public reaction and not as a verified fact.

Taken together, the GPT-6.1 Sol announcement is framed as an interim update within the GPT-6 family, designed to offer a cheaper, more accessible alternative to the top-of-the-range model (Astra), aimed both at ChatGPT users and at developers building applications and working with Codex. Given the nature of the material —a short promotional thread on social media— details common in more complete announcements fall outside its scope, such as price comparisons per uncached input token, context limits, the API availability date versus ChatGPT, or specific results on standardized tests.