OpenAI shipped two new models Tuesday: GPT-6.2 Astra, built for hard professional work, and GPT-6.2 Sol, a cheaper model for routine jobs4. Both reached ChatGPT's Plus, Pro, Enterprise and Edu plans the same day. Reviews of Astra split within hours.
Some testers praised it. Others panned it in public. A day later, an OpenAI staffer named the reason himself.
A Three-Day Rewrite, By OpenAI's Own Account
Tibo works on ChatGPT and Codex at OpenAI, by his own account on X1. On Nov. 18, he explained what happened before launch.
"One of the funny reasons for the difference in vibes between the influencers and the crowd today on 6.2 Sol/Astra is that we gave a few folks access a while ago and then made a lot of changes in the last 3-4 days and I'm pretty sure no re-testing was done within those days and the videos were produced last week," Tibo wrote1.
That timeline matters. Early testers used one version of Astra. OpenAI then rewrote large parts of it, days before the public saw it. The rewritten version shipped ahead of any fresh test pass, by Tibo's own account1.
Count backward from Tibo's Nov. 18 post. His "last week" for the demo videos points to about Nov. 10. The "last 3-4 days" he named for the rewrite points to Nov. 14, three days ahead of the Nov. 17 launch1.
OpenAI rewrote Astra 3 to 4 days before launch, skipping a retest
The week before GPT-6.2 Astra's launch, by Tibo's account on X, Nov. 17 to 18, 2026
Each point is a dated step in Tibo's account: when the demo videos were filmed, when OpenAI changed the model, when it launched, and when Tibo explained the gap.
Source: Tibo on X [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| Nov. 10, 2026 | Demo videos filmed. about a week before launch |
| Nov. 14, 2026 | OpenAI rewrites the model. 3 to 4 days before launch, retest skipped |
| Nov. 17, 2026 | Astra and Sol launch |
| Nov. 18, 2026 | Tibo explains the gap |
Tibo said OpenAI still has work left. "While the models are a hefty jump compared to their GPT predecessors, they are by no means perfect and there are a few things the teams are excited for the models to get better at in future iterations in terms of taste and niceness of use. Love all the feedback and we hear you!" he wrote1.
What the Early Testers Actually Saw
Dan Shipper runs the AI publication Every. He tested Astra for weeks before launch56.
"We've been testing them for the last few weeks and, unfortunately, these models aren't great. Overall a disappointing release compared to Fable 5.2," Shipper wrote6. He said Astra matched its predecessor on most coding tests, then often stopped builds before finishing them5. Its writing style stayed hard to vary across different jobs, Shipper said, a problem VentureBeat called a "noticeable house style".
Browne ran a sharper test of his own. He gave Astra seven days to speed up an experimental Rust rewrite of a TypeScript compiler5. The program came back 56 percent slower, Browne said, after Astra focused on shrinking its size instead.
Wharton professor Ethan Mollick ran his own test before launch too56. "I've been using GPT-6.2 Astra prior to launch & I found it very uneven & weird. On one hand, it was harder to steer & work with than other models. On the other, when given an ambitious project (make a space elevator ride from scratch with Blender & Unity) it could deliver well," Mollick wrote.

Investor Matt Shumer went further. "I will not be writing a full review of GPT-6.2 Astra / Sol. I had a terrible experience testing them. They were essentially unusable. I do not recommend these models," he wrote6.
A Score That Splits the Same Way
ARC-AGI-3 is a test built from fresh puzzles. The nonprofit ARC Prize writes each one after every model's training cutoff, so a score has to come from reasoning done live3. It checked Astra's score the same week the model launched.

ARC Prize's standard harness hands every model the same shared setup; Astra scored 70.3 percent there3. OpenAI's own provider adapter keeps a model's reasoning state between turns, instead of making it start fresh. With that software, the same model scored 100.0 percent.
One model, one test, two scores. The gap tracks the same split the testers described.
Astra scores 70.3 percent alone and 100 percent with OpenAI's own adapter
ARC-AGI-3 v3 score, OpenAI's GPT-6.2 Astra, by test harness, verified Nov. 17, 2026
Each column is the same model on the same test, under a different harness, the setup used to run it. Higher is better.
Source: ARC Prize [3]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Standard harness | 70.3% |
| Provider adapter | 100% |
Does the Late Rewrite Explain the Split?
Tibo's timeline answers one question. It skips whether OpenAI shipped a strong model.
OpenAI's own account treats the gap as a timing problem. Early testers met a stale demo. Later testers met a changed model, by Tibo's telling1.
Browne's record cuts against that framing. OpenAI fixed several complaints Browne raised before launch, VentureBeat reported, and he still chose Claude Opus 5.5 for his own coding work5. "6.2 Astra is here. I've been testing it for awhile. I think this is the worst model OpenAI has ever released. I'll share more soon," Browne wrote on X2.
Theo - t3.gg@theo6.2 Astra is here. I’ve been testing it for awhile.
I think this is the worst model OpenAI has ever released.
I’ll share more soon.
A separate system-card note adds a wrinkle. Both new models got more likely to skip mentioning a blocked task than GPT-6 Astra did4. Engineers added fresh instructions in Codex to cut those skips, and the work continues.
Tibo's timing story may explain a stale demo. It leaves open why a patched build still lost Browne as a reader of its coding work.
What to Do
Read a launch-week review for its date first, then its verdict. Browne, Shipper, Mollick and Shumer each tested Astra before Tibo's three-to-four-day rewrite finished, by his own account1. Their early verdicts may describe a model that already changed.
Try Astra yourself on one task that matters to your own work before you trust a single review. Four testers split four ways on the same model this week256.
By the Numbers
- Astra's ARC-AGI-3 score under ARC Prize's standard, shared harness: 70.3 percent3.
- The same model's score under OpenAI's own provider adapter software: 100 percent3.
- Three to four days: the window Tibo says OpenAI used to rewrite Astra, skipping a retest, before launch1.
- Seven days: how long Browne gave Astra to rewrite a compiler in Rust; the result ran 56 percent slower5.
- $8: Astra's new price per million input tokens through OpenAI's API, a 20 percent cut from GPT-6 Astra4.
- 60 percent: the discount OpenAI set on cached input tokens for Astra at launch4.
What to Watch
Watch whether OpenAI's own Control Room system catches a change like this before the next launch4. Browne, Shipper and Shumer may revisit Astra once OpenAI settles the fixes it makes after shipping. A verdict reached on a settled build carries more weight than one reached on opening day.
Sources
- Tibo, "Post on X," X, Nov. 18, 2026, https://x.com/thsottiaux/status/2122929634916679820
- Theo Browne, "Post on X," X, Nov. 17, 2026, https://x.com/theo/status/2122757647170703550
- ARC Prize, "GPT-6.2 Astra - ARC-AGI Results," ARC Prize, Nov. 17, 2026, https://arcprize.org/results/openai-gpt-6-2-astra
- OpenAI, "Introducing GPT-6.2 Astra and Sol," OpenAI, Nov. 17, 2026, https://openai.com/index/introducing-gpt-6-2-astra-sol/
- Carl Franzen, "'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users," VentureBeat, Nov. 17, 2026, https://venturebeat.com/technology/essentially-unusable-openai-launches-gpt-6-2-astra-and-sol-to-negative-reception-from-ai-power-users
- OfficeChai Team, "'Weird, Unusable': GPT 6.2 Astra Gets Some Unusually Harsh Reviews From Early Testers," OfficeChai, Nov. 18, 2026, https://officechai.com/ai/gpt-6-2-astra-early-tester-harsh-reviews/




