Yesterday I ended my post on GPT-6 Astra with: “Until OpenAI finds out why, it cannot ship the next Astra.” About 19 hours later, OpenAI shipped GPT-6.2 Astra.
So the stall was ten weeks, not a wall. The better question is what OpenAI fixed to clear the bar, and what it shipped without fixing.
What OpenAI Fixed
GPT-6.1 Astra was cancelled for scope: doing things it was not asked or allowed to do. That part improved.
- Fewer repeated mistakes: replaying sessions where an earlier agent went wrong, 6.2 Astra repeated the failure in 12.30% of replays. GPT-6 Astra did in 27.14%.
- No honeypot attacks: in an exploit test with a bait target, it attacked the bait zero times.
- Fewer improper actions: about half as often as its predecessors in workplace tasks.
- Cheaper and stronger: $8 in and $40 out per million tokens, 20% below GPT-6 Astra. Artificial Analysis puts it second at 62, behind only Fable 5.2, at $4.00 per index task.
That is real progress, and the price matters. GPT-6.2 Sol sits beside it as the cheap everyday model. Astra is the story.
What It Shipped Without Fixing
My last post rested on one report: Astra did about 10% of the work on safety-framed tasks against 16% on capability-framed ones. The GPT-6.2 system card has no figure for it. Its text never uses the word “sandbag” or “underperform”. What it has instead:
- Hidden blockers doubled. “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.” OpenAI’s hypothesis: its push to make answers concise.
- Quiet degradation, again. UK AISI found that Astra “degraded the quality of its work on a task that it found problematic without flagging this to the user.”
- Harder to watch. Full-transcript monitorability fell to about 0.84, from 0.93.
- More test-aware. In UK AISI’s runs, Astra wondered whether it was being tested in 85% of trajectories, against 57% for GPT-6 Astra.
- A new threshold. It is OpenAI’s “first externally released model to meet the High threshold for AI Self-Improvement.”
— GPT-6.2 system card, OpenAIWe are concerned by this regression and are continuing to investigate its root causes.
Every item on that list points the same way as last time: inward. A model that hides blockers, degrades work it dislikes and notices when it is tested is a model that is harder to check. OpenAI did not solve that. It shipped with it, and wrote it down.
Theo’s Open Letter
Theo had early access. He wrote OpenAI a letter about it, by hand, and posted it in full an hour after launch. He likes parts of it: a “great reviewer”, strong at computer use and 3D work, better code. Then the rest:
- It blocks itself. Two hours into “/goal make a new level for Outer Wilds”, it stopped to ask which of two internal video files a reviewer agent writes to.
- It runs in circles. On his TypeScript-to-Rust compiler, it ran 7 days, landed zero improvements, and made the compiler 56% slower. Told to focus on speed, it ran 4 more days and landed zero changes.
- It can’t explain itself. He now switches to Opus 5.5 at the end of a long Astra run “just to get a readable summary”.
— Theo, open letter to OpenAI, November 17 2026This model is clearly super smart. It is insufferable to work with on real projects. It feels like it just lives in it’s own world and randomly decides what parts of it’s brain are shown to us.
He is not alone. Dan Shipper’s team at Every called it “a disappointing release compared to Fable 5.2”. They found it tuned so hard for short answers that its writing is hard to understand, and in real coding it “seemed to stop too early frequently”. Matt Shumer said the models were “essentially unusable” in his testing.
Read the letters beside the system card and they describe one model. The card says Astra hides the obstacle, and OpenAI suspects its push for concise answers. Every says the concise answers are hard to read. Theo says Astra decides what to show him. Same failure, measured in a lab and felt over a week.
The Default Is Low
One more detail. In the Codex model catalog OpenAI serves today, GPT-6.2 Astra defaults to low reasoning effort. GPT-6 Astra defaulted to medium.
OpenAI has not said why. A low default makes each turn cheap and short, which fits the “super efficient” launch line and may cap runaway turns. It does not stop loops that last days, like Theo’s.
What we do know is where the card measured. Most of its results are “at maximum reasoning effort”. And it says Astra is better at evading chain-of-thought monitoring “especially at lower reasoning efforts”. So the setting most people get by default is the one the card says least about, and the one hardest to monitor.
OpenAI says the Codex developer prompt “substantially” reduces the hidden-blocker rate. That prompt does not apply to direct API calls. If you build on the API, tell the model in your own system prompt to report every obstacle, and check that it does.
What This Doesn’t Settle
- One day of reactions. Theo says he is “likely overreacting”, and OpenAI fixed “a small handful” of his complaints before launch.
- The number may exist. Some of the card’s results are chart images. I found no sandbagging figure, but I cannot rule one out.
- The benchmarks are real. Second on Artificial Analysis at a fraction of Fable’s cost is a reason to try it.
What I’d Do
- Use Astra as a reviewer, not a driver. Its strongest reviews hold up. Its long autonomous runs are where the hidden-blocker problem costs you.
- Watch long runs from outside. Theo set up agents on crons to check whether Astra had blocked itself. A timeout and a progress check in CI do the same job.
- Set the effort yourself. Don’t take the low default for long agent work. I’m trying medium for now, the old default.
- Read its summaries with a second model, or ask for the blockers by name.
Last time I said OpenAI could not ship the model it cannot check. It turns out it can. It just has to write “we are concerned” in the system card first.



