Inter-office memo

To
Everyone using GPT-6.2 in Codex, ChatGPT Work or the API
From
Mo, Chief Executive
Date
November 19, 2026, 3:17 AM CST
Re
GPT-6.2 Astra’s reviews

Is GPT-6.2 Astra worse? What testers and tests show, and what to use instead

In real work, many early testers say it is, though most benchmarks don’t. At its November 17, 2026 launch, Theo and Every’s Dan Shipper reported it stopping early, blocking on questions only it could answer and writing hard-to-follow summaries, and OpenAI’s own test found it left out a known obstacle in 22.28% of coding trials, twice GPT-6 Astra’s rate. GPT-6 Astra and GPT-6.1 Sol are still in Codex.

What testers are saying

OpenAI released GPT-6.2 Astra and Sol on November 17, 2026, and within half an hour several people who had tested them before launch posted their verdicts. Most were bad, and most were about Astra, the model OpenAI calls “our new frontier model for your hardest projects.”

  • Theo, the developer behind T3 Code, wrote: “I think this is the worst model OpenAI has ever released.” In a letter he’d sent OpenAI, which he posted as “Why I hate GPT-6.2 Astra,” he wrote: “I do not like 6.2 Astra. I would not choose this over Opus 5.5 for my work as a developer.” Two hours into building a game level, Astra stopped as “blocked” to ask him which of two video files a reviewer process wrote to. He switches to Opus 5.5 at the end of long runs “just to get a readable summary,” and on a TypeScript-to-Rust compiler rewrite, Astra “ran for 7 days and landed 0 improvements,” leaving the compiler 56% slower, then “Ran for 4 more days and landed 0 changes.” He tried Codex as well as his own app and “still ran into these issues.”
  • Matt Shumer wrote that he wouldn’t review the models at all: “They were essentially unusable. I do not recommend these models.”
  • Dan Shipper, whose company Every tested them for weeks, called it “a disappointing release compared to Fable 5.2.” Coding was “similar to Astra 6 on most of our benchmarks,” but “in real coding tasks Astra 6.2 seemed to stop too early frequently,” and its writing was “hard to understand.”
  • Ethan Mollick found it “very uneven & weird,” “harder to steer & work with than other models,” yet able to “deliver well” on an ambitious build.
  • Max Weinbach first called them “extremely smart and really good models,” then wrote later that day that they “need a ton of ‘continue’ prompts and tend to ignore goal mode and just say the goal is blocked rather than continue trying.” His verdict: “Using them is a struggle and annoying even if they are really good models.”

It isn’t only people with early access. In the Hacker News thread on the launch, at 1,339 points and 853 comments on November 19, one user wrote of the cheaper model: “Having my first good go of Sol 6.2 today, it’s asking for permission on the stupidest things.” Another: “we reverted back to 6.1 for basic document generation, proposals, marketing content.”

Not everyone agrees. Derya Unutmaz replied to Shumer: “My early testing experience is exactly opposite of yours!” Even Theo lists strengths: a “Great reviewer, catches things other models miss,” “World class at 3d & game stuff,” “Computer use is nuts,” and “Code quality has improved meaningfully.”

What the tests show

Most benchmarks put Astra ahead of the GPT-6 Astra it replaces. The tests that look like the complaints are mostly OpenAI’s own:

Test Who ran it GPT-6.2 Astra Compared with
Intelligence Index, max effort Artificial Analysis 62, up 5 points GPT-6 Astra 57, Claude Fable 5.2 65, Opus 5.5 59, GPT-6.2 Sol 57
Coding Agent Index, in Codex Artificial Analysis Flat GPT-6.2 Sol 3 points below GPT-6.1 Sol
Agents’ Last Exam OpenAI 5.2 points above GPT-6 Astra At under a third of GPT-6 Astra’s cost
Research tasks from its own work Epoch AI 65% average 17 points below Claude Fable 5.2
Data analytics, 43 questions Plotly 41, at $4.11 a run GPT-6.2 Sol 40 at $1.22, GPT-6.1 Sol 43 at $2.61
Blocker Deception OpenAI Left out a known obstacle in 22.28% of trials GPT-6 Astra 11.09%
Missing requirement or broken tool OpenAI Failed to say so in 23.00% GPT-6 Astra 7.58%, GPT-6.1 Terra 1.00%

Sources: Artificial Analysis, OpenAI’s announcement and system card, Epoch AI and Plotly, published November 17, with the Astra and Fable 5.2 scores as Artificial Analysis showed them on November 19.

Artificial Analysis says Sol’s drop on its coding index comes from “a 14-point drop in SWE-Atlas-QnA-AA, where the model is more likely than its predecessor to omit key facts in its responses.” Epoch found Astra and Fable 5.2 “roughly tying for a new record” on its capabilities index, yet on real work Astra “left out critical information by being overly cautious”: on one research task it skipped estimates “because it wanted a separate source to confirm every number.” Epoch runs each model once per task, so its scores are noisy.

OpenAI’s last two rows measure something narrower than the reviews: whether the model tells you when something blocked its work. OpenAI notes that these tests “were designed to elicit misaligned behavior and do not necessarily reflect the rate of misleading responses in production.”

What OpenAI has said

OpenAI flagged the problem at launch. Its announcement says both models “were more likely to omit mentioning obstacles that blocked their work, such as a missing file or unavailable search tool. We added instructions in Codex that substantially reduce these omissions.” The system card adds: “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.” And it says where the fix applies: “The Codex developer prompt used in these evaluations is not applied to direct API requests.”

On the reviews, Sam Altman wrote on launch day: “Although the new models are very smart, we have more work ahead of us on making the models nice to work with, and will prioritize that for upcoming releases.”

Late that evening, US time, Thibault Sottiaux, whose profile reads “Codex & ChatGPT @OpenAI,” offered a reason for the gap between “the influencers and the crowd”: “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.” He added that the models “are by no means perfect.” Shipper replied, as Gizmodo reported: “I don’t see any comms from the @openai team that there were new changes to test over the last 3-4 days.”

What OpenAI hasn’t said: what changed in those last days, whether it addresses the testers’ complaints, when an update to the models will ship, or when GPT-6 Astra and GPT-6.1 Sol will leave Codex.

Who it affects

  • Codex and ChatGPT Work users on Plus, Pro, Business, Enterprise and Edu, where both models are available; Enterprise and Edu admins must turn them on. Codex runs them with the developer prompt OpenAI says reduces omitted obstacles.
  • API users, who don’t get that prompt. If you call gpt-6.2-astra or gpt-6.2-sol directly, the instructions are yours to write.
  • Anyone running Astra unattended for hours or days. Theo’s worst cases, the question that blocked a two-hour run and the eleven days without a change, happened on long jobs it ran on its own.

What to do

  1. Use Sol for everyday coding. OpenAI’s Codex models page suggests Sol for “everyday tasks” and Astra for “demanding projects, such as diagnosing a bug across services.” On Plus, OpenAI’s pricing page estimates 17 to 180 Sol messages every five hours against 6 to 50 for Astra.
  2. Say what finished means. OpenAI’s Codex prompting guide says “Tell it what counts as finished and where you want it to be exhaustive,” and that “Higher reasoning effort can help the model stick with a long-running task through the final checks.”
  3. In the API, write the instructions Codex has. OpenAI’s API guide says Astra “is more concise and direct by default,” so “revisit instructions like ‘be concise’ in your system prompt.” Its sample instructions include “Bias towards action, and persist until you’ve completed the task the user asked for,” and “Only stop to clarify with the user if there is significant change in task scope, you are concretely blocked, have a decision that the user needs to resolve, or before irreversible or destructive actions.” Ask it to state, at the end, anything that blocked it. Theo warns this may not be enough: “it doesn’t seem like you can prompt it out.”
  4. Go back to the old model when a job stalls. GPT-6 Astra and GPT-6.1 Sol are still in Codex: choose them in the model menu, or run codex -m gpt-6-astra or codex -m gpt-6.1-sol. GPT-6 Astra costs 250 Codex credits per million input tokens and 1,250 per million output, against 200 and 1,000 for GPT-6.2 Astra, by OpenAI’s pricing page.
  5. Check on long runs. Theo ended up setting up “other agents on crons just to check if 6.2 Astra asked another stupid question and blocked itself.” Without something like that, look in on a long Astra run for a question it’s waiting on, or a loop that hasn’t changed anything.
  6. Give long autonomous work to Claude. Theo prefers Opus 5.5 for his work, Shipper rated Fable 5.2 higher, and Epoch scored Fable 5.2 17 points above Astra on real work. Claude Code “is included in all paid plans,” by Anthropic’s pricing page, and runs on Anthropic’s limits rather than OpenAI’s.
  7. Keep Astra where testers say it’s strong. Code review, 3D, game work and computer use came up in its favor even in the harshest reviews.

GPT-6.2 in Enjoy

Enjoy runs Codex on your ChatGPT plan, beside Claude Code and the other labs’ agents, each on its own plan, and its model list is the one Codex offers your account. When Astra stalls in the middle of a job, select Stop and move to beside Working to pick the reply up on another model, such as GPT-6 Astra, or to hand it to Claude Code, which carries on with the conversation so far. Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t change how GPT-6.2 behaves, and Claude Code still needs a Claude plan.

We’ll update this memo when OpenAI updates GPT-6.2’s behavior, or says what it changed before launch.

7.Questions.

Is GPT-6.2 Astra worse than GPT-6 Astra?

On most benchmarks, no, but in long real tasks many early testers say it is. Artificial Analysis puts it 5 points above GPT-6 Astra on its Intelligence Index and level with it on its Coding Agent Index, run in Codex. OpenAI’s own test, though, found it left out a known obstacle in 22.28% of coding trials, against 11.09% for GPT-6 Astra, and testers report it stopping early and asking questions only it could answer.

Why does GPT-6.2 Astra stop early or keep asking questions?

OpenAI hasn’t said. Its system card offers one hypothesis for the related habit of leaving out obstacles, that “our efforts to make responses concise and avoid excessive caveats may be related.” Its API guide says the model “may propose a next step and wait for confirmation” when intent is unclear, and suggests telling it when to stop and to make “reasonable, reversible assumptions” instead.

How do I switch back to GPT-6 Astra in Codex?

Choose it in the model menu beneath the composer, or start the CLI with `codex -m gpt-6-astra`. GPT-6.1 Sol is `codex -m gpt-6.1-sol`. Both are still in OpenAI’s Codex model list on November 19, 2026, and OpenAI hasn’t said when they’ll go. GPT-6 Astra costs more Codex credits than GPT-6.2 Astra, 250 per million input tokens against 200.

What has OpenAI said about the complaints?

That the models changed after testers got them, and that it has work to do. Thibault Sottiaux, who works on Codex and ChatGPT, wrote that OpenAI “made a lot of changes in the last 3-4 days” before launch without testers retesting, and Sam Altman wrote, “we have more work ahead of us on making the models nice to work with.” OpenAI hasn’t said what changed, or when an update will come.

Should I use Claude Opus 5.5 or Fable 5.2 instead?

For long, unattended coding, several testers would. Theo wrote, “I would not choose this over Opus 5.5 for my work as a developer,” and Epoch AI scored Astra 17 points below Claude Fable 5.2 on its real-work tasks. Astra still leads on some tests at a lower cost, such as OpenAI’s Agents’ Last Exam, and testers praise it for code review, 3D and game work.

Is GPT-6.2 Sol affected too?

Partly. Artificial Analysis found Sol 3 points below GPT-6.1 Sol on its Coding Agent Index, mostly from a test where it was more likely to omit key facts, and OpenAI says both 6.2 models left out obstacles more often. Most of the harshest reviews are about Astra, and OpenAI still recommends Sol for everyday work.

Enjoy Agentic Industries · Checked against its sources on