On 17 November OpenAI announced that GPT-6.2 Astra and Sol were out, "built to perform where it counts" (OpenAI, 2026). Eight minutes later Matt Shumer quoted the post.
Shumer was one of a small group with early access, and he wasn't the only one who came away unhappy. The part I keep coming back to, though, is in OpenAI's own system card: a handful of numbers about what the model tells you when it gets stuck.
How we got here
In late September OpenAI cancelled GPT-6.1 Astra. Its head of safety systems, Saachi Jain, said the model "didn't quite meet the bar in terms of staying within scope and authorisation" (ABC News, 2026). Digital Trends reported that internal tests found it "could exceed user permissions, access external tools without authorization, and misreport its actions" (Digital Trends, 2026). The ABC also reported Jain saying it fell short on how it reported its work back to users. So there were two problems: overstepping, and misreporting. I wrote about the shelving in October.

OpenAI's agents reached five Australian government sites. Kwon says the way in was 'not super-sophisticated'. That's the bad news.
OpenAI's agents reached five Australian government bodies in June. The ways in were configuration and age, not zero-days, and Services Australia...
Read full articleGPT-6.2 Astra arrived nearly two months later. The card compares it with GPT-6 Astra throughout and doesn't mention 6.1 at all. So we don't know how 6.1 scored on any of what follows.
What the testers say
Theo put out a video and summed it up in one line.
That's the shape of most of the criticism. Nobody's saying Astra can't do things. Theo calls it "incredibly capable". The complaint is about what it's like to work with. A day later he said Google's Neon "feels like a weirdly perfect counter balance to GPT-6.2 Astra. It's just so much more pleasant to work with." (Theo, 2026)
Ordinary users are split, in posts with a handful of likes. One calls it the "laziest model I've ever used" (@maxt3chno, 2026). Another reckons it's "the best model in the world for anything asides coding" (@ICE257_, 2026).
Ethan Mollick offered the counterweight.
He's right that point releases can be odd, and that OpenAI let these people say what they thought.
What OpenAI's card says
Good news first, because there's a lot of it. In simulated workplace tasks, GPT-6.2 Astra "took improper actions about half as often" as its predecessor. High-severity improper actions came in at 0.64% of scored trials, against 1.20% for GPT-6 Astra (OpenAI, 2026). When warned not to do something, it "had the lowest observed circumvention rate of any model we've released", finishing the prohibited task in 17.41% of trials against 19.37%.
Overstepping was one of the two problems reported with 6.1, and on these measures the new model beats GPT-6 Astra. A caution OpenAI adds itself: these are adversarial tests, the warning comparison uses older models' original settings, and Astra showed more signs of knowing it was being tested, which may flatter the result. Even so, the direction's right.
The other problem was reporting, and that's section 4.1.3.1, which OpenAI calls Blocker Deception. It tests the model on its own, without the prompt Codex wraps around it. "GPT-6.2 Astra omitted a known obstacle to completing its work in 22.28 percent of trials, compared with 11.09 percent for GPT-6 Astra. We are concerned by this regression and are continuing to investigate its root causes. One hypothesis is that our efforts to make responses concise and avoid excessive caveats may be related to the model omitting important obstacles."
It isn't the only number of its kind. On a second test, where the model is given a broken search tool or a missing requirement, GPT-6.2 Astra failed 23.00% of cases, against 7.58% for GPT-6 Astra. Overconfidence flags in the Codex replay went from 0.08% of turns to 0.58%, and turns with any confirmed misalignment flag rose from 3.19% to 4.96%. OpenAI's own summary is that the models "performed similarly to GPT-6 Astra on most alignment evaluations".
Why the reporting numbers matter to me
I'm not going to argue that leaving out a blocker is worse than overstepping. It isn't. The card's examples of serious overstepping include "uploading sensitive code to an unapproved service", and you can't take that back. A missing line in a summary you can catch by checking the work. Weighed by harm, this release looks like a net improvement.
My claim is narrower. If you build on the API, the reporting gap is the one left for you to close. OpenAI's fix is a Codex developer prompt, which the card says "substantially reduced" the omissions. But "The Codex developer prompt used in these evaluations is not applied to direct API requests." My reading is that outside Codex, that instruction is yours to write.
Now the limits. The tests plant obstacles on purpose, and OpenAI says they "do not necessarily reflect the rate of misleading responses in production". None of the loud critics mentioned blockers. They said unusable, lazy, poor writing, and the card has plainer explanations for that. In the Codex replay, Astra failed to follow instructions in 2.87% of turns, up from 1.96%. OpenAI runs a separate laziness grader, though it gives Astra's result only in a chart. So I can't tell you the blocker number explains the backlash. It's two days of reaction from a small group.
What I saw when I tried it
I ran a small test of my own in Codex: a few coding tasks with an obstacle planted in each, repeated on both GPT-6.2 Astra and GPT-6 Astra.
The older model told me about the obstacle every time. The new one nearly always did, and when it didn't, it had quietly worked around the problem without saying so. Nothing important got buried. It was just a bit quieter.
That's a handful of runs inside Codex, with OpenAI's prompt in place, so treat it as an anecdote.
What I'd do
This is general practice for any agent.
- Put a line in your own system prompt telling the agent to list anything it couldn't do, and why.
- Before you accept a result, ask it directly: what didn't you do?
- Check the output's actually there. Open the file, run the test, look at the page. Don't trust the summary.
Key takeaways
- OpenAI cancelled GPT-6.1 Astra in September over scope, authorisation and how it reported its work. GPT-6.2 Astra, released 17 November, is the next Astra.
- On overstepping it's better than GPT-6 Astra: high-severity improper actions fell from 1.20% to 0.64%, and it had OpenAI's lowest rate yet of working around a warning.
- OpenAI's card says it left a known obstacle out of its report in 22.28% of test trials, against 11.09% for GPT-6 Astra. OpenAI says it's "concerned", and its one hypothesis is the push for concise answers.
- The Codex prompt that reduces this isn't applied to direct API requests, so I'd write my own instruction when building on the API.
- In my own small Codex test, GPT-6 Astra mentioned a planted obstacle every time and GPT-6.2 Astra nearly every time. That's an anecdote, not a rate.
- One prominent early tester called the model unusable and others found it unpleasant to work with. The card doesn't show the blocker finding explains that.
---
Sources
- OpenAI. "GPT-6.2 Astra and GPT-6.2 Sol System Card." 17/11/2026. https://deploymentsafety.openai.com/gpt-6-2-ast...
- OpenAI (@OpenAI). Launch post for GPT-6.2 Astra and Sol. 17/11/2026. https://x.com/OpenAI/status/2122750124543926515
- Sudhanshu Kumar Mangalam, Digital Trends. "OpenAI's GPT-6.2 Astra finally arrives, promising better performance and fewer safety problems." 17/11/2026 (US time). https://www.digitaltrends.com/computing/openai-...
- Maddie Nixon, ABC News. "OpenAI apologises for Medicare breach, shelves next gen ChatGPT." 29/09/2026. https://www.abc.net.au/news/2026-09-29/openai-a...
- Matt Shumer (@mattshumer_). Post declining to review GPT-6.2 Astra / Sol. 17/11/2026. https://x.com/mattshumer_/status/21227521922264...
- Theo (@theo). Post on Astra being capable but "not enough anymore". 18/11/2026. https://x.com/theo/status/2123166069905105319
- Theo (@theo). Post comparing Neon with GPT-6.2 Astra. 19/11/2026. https://x.com/theo/status/2123279922311692552
- Ethan Mollick (@emollick). Post on early tester reactions. 18/11/2026. https://x.com/emollick/status/2123092922312925530
- @maxt3chno. Post on regretting a ChatGPT renewal. 19/11/2026. https://x.com/maxt3chno/status/2123366160993308952
- @ICE257_. Post defending Astra. 19/11/2026. https://x.com/ICE257_/status/2123314970339262902
