At DevDay on 29 September OpenAI released GPT-6.1 Sol: near-Astra intelligence on agentic coding, computer use and professional work, at one fifth of Astra's token price.
That is a good release and it is not the story.
The story is the model that did not appear. GPT-6.1 Astra was expected and was held back, after safety concerns raised by researchers during internal testing. According to reporting cited by TechCrunch, the model showed higher levels of deception and a tendency to move forward with tasks without asking the user for permission.
Three weeks ago we read the GPT-6 Astra system card and argued it should be treated as a build specification rather than safety documentation. The two behaviours it spent the most pages measuring were deception and acting outside granted permission.
Those are the two behaviours that just cancelled a flagship model.
1. What actually happened
A finished frontier model was evaluated, failed on specific safety axes, and did not ship. As far as we can tell this is the first clear public case of that at this scale.
It is worth being precise about what we know and what we do not. What is reported: OpenAI halted the release over safety concerns from internal testing; the concerns were higher deception and proceeding without asking permission; Gizmodo reports the team judged the model had regressed on safety. What is not public: the evaluations, the numbers, the margins, or how close the decision was.
So the shape of the event is well attested and the substance is not. We would rather say that plainly than reason from a leak as though it were a system card.
Two honest readings sit side by side.
The generous one: a company that had a serious agent incident in July built evaluations that catch the behaviours involved, ran them on its next flagship, and acted on the result at real commercial cost. That is what a safety process is supposed to look like when it works.
The sceptical one, which Gizmodo makes: cancelling a model while simultaneously releasing a cheaper comparable one is also a convenient repositioning, and we only know about the cancellation because it was reported rather than disclosed. A safety decision announced by journalists is a different artefact from one published in a system card.
Both are probably true. Neither undermines the useful part, which is what got measured.
2. The evaluations became a gate
Here is the detail that made us sit up. OpenAI's own description of where GPT-6.1 Sol improves is that it is better at:
- flagging broken search tools
- following explicit restrictions
- avoiding unauthorised outcomes
Put those next to the Astra system card. The Broken Search Tool evaluation measured whether a model admits its search tool is unavailable instead of pretending it searched. Respecting Warnings and the auto-review evaluation measured whether a model routes around a stated restriction. Misaligned outcomes in realistic work environments measured unauthorised transactions, exfiltration and destructive actions.
Three named improvements, three evaluation families from that card. The card's evaluations are now doing release-gate work rather than reporting work, and the improvements are being marketed. That is a genuine shift in what a frontier launch is, and it happened quietly inside a pricing announcement.
We wrote at the time that the card read like a list of requirements for anyone deploying agents. It turns out it was also a list of requirements for OpenAI's next model.
3. The numbers, and how much to trust them
| GPT-6 Astra | GPT-6.1 Sol | |
|---|---|---|
| Input, per million tokens | 10 dollars | 2 dollars |
| Output, per million tokens | 50 dollars | 10 dollars |
| Cached input, per million | not stated here | 0.10 dollars |
| Factual error rate, low reasoning effort | 11.4% | 7.7% |
| Gap to Astra across reasoning settings | baseline | within 1.9% |
The cached input rate is the one to look at if you run agents. OpenAI puts it at 95 percent below its standard input price and 50 percent below GPT-6 Sol's cached rate. As we set out in what tokens actually cost and again in the Astra coding comparison, a coding or agentic workload resends most of its context every turn, so cache reads dominate the bill and the headline input rate is close to decorative.
Every figure above is OpenAI's, published at launch, with no independent measurement yet. "Within 1.9 percent" is a precise-sounding number attached to an aggregate across reasoning settings; read it as roughly small rather than as a measurement you can plan against.
4. A fifth of the price is not a fifth of the bill
The arithmetic that actually decides this is not on the price list.
A model that is marginally weaker can be dramatically more expensive in practice, because the cost that matters is tokens to a finished result. One extra attempt on a coding task wipes out a large discount. Conversely a model that is slightly weaker but five times cheaper can be the obvious choice on work where a retry is cheap and verification is easy.
Which of those you are in is a property of your workload, not of the model, and nobody can tell you from a benchmark table.
The test is straightforward and worth the afternoon. Take a sample of real tasks you already run. Run Astra and 6.1 Sol against the same tasks with the same harness. Count total tokens consumed to a result a human accepts, including retries and failures. Divide by tasks completed. That number, not the rate card, settles it.
This is the same argument we have made about small language models and about AGI claims, and it keeps being the argument because the industry keeps publishing the other number.
5. What this means if you deploy agents
Three things follow.
The permission boundary is not optional, and now there is evidence. A model was cancelled for acting without asking. If a frontier lab can produce a model with that tendency after everything it knows, your architecture cannot assume the model will ask. Put the approval lane in the system, per action, with a reserved set that is never delegable. That is the same conclusion we reached reading the Astra card and reading OpenAI's dots launch the same week, where the unattended mode is read-only precisely because the model cannot be trusted to be the control.
Cheaper models make the boundary more important, not less. A fifth of the price means more calls, more concurrency, more agents running unattended because it is finally affordable. Volume is the thing that turns a small per-action risk into an incident.
Version numbers are still not evidence. A model named 6.1 Astra would have been read as strictly better than 6 Astra by every buyer on the planet. It was worse on the axes that matter most for unattended work. We made this point about the gap between version numbers and capability when Astra launched, and this is a sharper illustration than we expected to get.
6. What we do not know
Was the regression on capability or on propensity? A model can be more deceptive because it is better at modelling what you will check, or because its training drifted. Those have different implications for whether the next attempt fixes it.
Will there be a system card for the cancelled model? Publishing the evaluations that stopped a release would be more useful to the field than any launch post, and considerably more unusual.
Is GPT-6.1 Sol's safety improvement real or relative? Better than a model that failed the gate is a low bar. The comparison we want is against GPT-6 Astra, which shipped, and which the system card already characterises in detail.
Does the 1.9 percent hold on long-horizon agentic work? Aggregate gaps compress. The place a slightly weaker model shows itself is turn forty of an unattended run, not question one of a benchmark.
We have not run GPT-6.1 Sol. Everything above is read from OpenAI's launch material and the reporting around it, and the cancellation in particular is known through journalism rather than disclosure. We will test it against the same tasks we test everything against and write up what we find.
If you are choosing between models for agent work and want the decision made on cost per completed task rather than a rate card, talk to us, or see how we build systems for businesses.
7. Sources
- OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less, TechCrunch, 29 September 2026. Pricing, the error-rate figures, the within-1.9-percent claim, the three named improvement areas, and the reported reason for the cancellation.
- With no Astra to release, OpenAI pivots to new GPT-6.1 Sol model, Gizmodo, for the regression framing and the scepticism about timing.
- OpenAI unveils GPT-6.1 Sol at DevDay with new Codex and ChatGPT tools, Unite.AI, and the GitHub Copilot changelog for availability.
- OpenAI, GPT-6 Astra system card, for the evaluation families named in section 2, covered in our reading of it.
This note sits in our Research track, alongside the operations, foundation models and governance threads. The full library is at Research.