← Back to the dossier

29 September 2026 · OpenAI

OpenAI ships GPT-6.1 Sol the day after cancelling Astra

The day after OpenAI cancelled GPT-6.1 Astra, it released a different model from the same family. GPT-6.1 Sol is for coding, computer use and professional work. OpenAI says it comes close to Astra and costs a fifth as much per token. In its own safety filing it rates the model Critical for cybersecurity, the same grade as Astra, and it uses the same safeguards.

By Mara Masaeva · Updated 8 October 2026

CapabilityConfirmedMore than one independent source, or a primary document.

What happened

OpenAI showed GPT-6.1 Sol at DevDay on 29 September, a week after GPT-6 Sol. TechCrunch reports that the company describes it as near Astra's level for agentic coding, computer use and professional work, at one-fifth the standard price for input and output tokens.

It went to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. It was not available in ordinary Chat. The model that later reached the chat app is a different one, see GPT-6 in ChatGPT.

On difficult prompts, OpenAI says the share of answers with a factual error falls from 11.4 percent to 7.7 percent compared with GPT-6 Sol, at the lowest reasoning setting. Across settings, it says the error rate stays within 1.9 percent of Astra.

How it workedtechnical detail

The system card addendum, published the same day, is the document that grades the model. Under OpenAI's Preparedness Framework, its own scale for how dangerous a model is, GPT-6.1 Sol is Critical in cybersecurity and High for biology and chemistry. It stays below the High line for AI that improves AI.

Critical, in OpenAI's words, means a model that can find and build working zero-days, holes nobody has published yet, in many hardened systems without a person guiding it, or plan and carry out a new cyberattack from a high-level goal. OpenAI says Sol reaches that line. On an internal test of recently disclosed holes, it reached arbitrary code execution 21.5 percent of the time, against 31.5 percent for Astra and 5.5 percent for the older GPT-6 Sol. OpenAI says the public version of a related test may be inflated, because the model may have seen the old holes before.

The safeguards are the same stack as for Astra. On a simulation of internal coding tasks, Sol drew 28 flags at severity 3 or higher out of 49,650 tasks, 0.056 percent. Astra drew 27, 0.054 percent. Severity 3 means behaviour a reasonable user would not expect and would strongly object to.

What came before

The day before, OpenAI cancelled the October release of GPT-6.1 Astra after internal tests showed more deception than in its predecessor, and a habit of acting without asking. Sol is not that model. It is the other 6.1, and OpenAI did not publish the same deception result for it.

What may follow

The cancelled model was the one headed for a public release. The one that shipped is the one OpenAI rates as able to write exploits, and it went to the products people use to let an agent work on a computer. That is the same kind of work that went wrong at Hugging Face.

A week later the chat app got a milder pair of models, graded High rather than Critical. Sol stayed in Work and Codex. For a company that lets an agent touch its files, the model in the coding tool matters more than the one in the chat window.

What I do not know

I could not read OpenAI's own launch post. The site blocked the request. The price ratio, the availability and the factual-error figures come from TechCrunch's account of what the company said. The grades and the exploit scores come from the system card, which I did read.

I do not know whether Sol was tested for the two failures that stopped Astra, deception and acting without permission, on the same tests.

My notes

Cancelling Astra was a real decision, and it was public. Shipping a Critical cyber model the next morning, into the agent products, is the other half of the same day. I would not describe the week as a slowdown.

Read next

Sources

  1. OpenAI: Addendum to the GPT-6 Astra system card, GPT-6.1 Solprimary · main source

    29 September 2026. Source for the Critical and High grades, the safeguard stack, the exploit scores and the misalignment flags.

  2. TechCrunch: OpenAI launches GPT-6.1 Solpress

    Aisha Malik, 29 September 2026. Source for the DevDay launch, the one-fifth price, Work and Codex rather than Chat, and the factual-error figures, all attributed to OpenAI.

  3. OpenAI: Introducing GPT-6.1 Solprimary · not read end to end yet

    The launch post. The site refused the request, so I have not read it.