GPT-6.2 Launch Day: Real Feedback on X and Its Limits, Read in Five Groups
On day one, one of Matt Shumer's posts was viewed about 1 million times, and it said the model was "basically unusable". Theo's long post drew about 1.67 million views. The same day, Artificial Analysis (AA) ranked Astra second on its Intelligence Index.
The contrast is sharp and easily turned into a shouting match. This article reads it differently: it sorts the original posts by type of problem, notes for each the author, the conditions and the facts that can be checked, and then says what this feedback can and cannot show.
The X posts cited are ones we read in the original, or their publicly visible content. Claims whose original posts we did not read are marked with the relaying source, or not cited. This article is not a test run by Guishu Intelligence.
Background
OpenAI released GPT-6.2 Astra and Sol on November 17, 2026 (US time). According to ChatGPT Learn the entry points are ChatGPT Work, Codex and the API. Several posters had early access before release, and their assessments came within hours before and after launch, after anywhere from several weeks of use to much less.
The data on these posts (views, likes) is the public count at the time of reading. It shows only the scale of the discussion and does not show that an opinion is correct.
Group 1: Hard to use, especially under heavy or long-horizon use
Matt Shumer (@mattshumer_). His first post says he would not write a full review because the testing experience was bad, the model was basically unusable and he does not recommend it; at the time of reading it had about 5,500 likes and about 1 million views. A follow-up adds that he had never written a negative review before and this was an exception; light use is fine, but heavy use "doesn't work at all". Another post says he mainly uses Fable 5.2 and some Opus 5.5 at the moment, and when he tried Astra as a sub-agent, Opus or Fable got frustrated and redid the work.
What can be checked: he had early access, and his main tools are Anthropic models. What cannot: his specific tasks and how they failed, which the posts do not give.
Theo (@theo). He calls it the worst model OpenAI has released. His long post gives concrete cases: during long runs it asks internal questions that cannot be answered; its explanations are hard to read, and he now uses Opus to summarize Astra's output; a ts-rust project ran seven days with zero improvement and a compiler 56% slower, then four more days with zero changes. He also says he may be overreacting. His token-spending history is his own account and has not been independently verified.
Dan Shipper (@danshipper). Every tested for several weeks and found it disappointing compared with Fable 5.2. The long text of the original X post was truncated when we read it, so the details below come from VentureBeat and Implicator: writing is worse and too terse; end-to-end builds often stop too early; the default design shows instincts but a clear house style. Every's in-depth review is behind a paywall, and we read only the free opening. It says the model can handle certain limited tasks that were previously unsolved, but the predecessor was already doing well and overall it "leaves us wanting more".
Group 2: Highlights, but harder to steer
Ethan Mollick (@emollick). He used Astra before release and says it is very uneven and a bit strange and harder to guide, but that a space-elevator experience built from scratch in Blender and Unity could turn out very well. In a follow-up: the world was playable in the browser with simulated zero gravity, but he found he would not pick the new model first because it needs more guidance and pushing. Later still he wrote that this is the first OpenAI frontier release since GPT-4.5 that is not an obvious step forward; he thinks it is strong on some complex work, but his experience and that of others is often frustrating.
Pietro Schirano (@skirano). After early testing he thinks the day-one reaction was unfair and that most problems were in the harness, which he had raised during testing; with it removed the model was "flying", and he then showed one-shot 3D and game examples. This is a personal judgment with no publicly reproducible comparison.
Chinese-language sources. Kazike and AIZ Xiaozhu's WeChat article sums it up as "lopsided" (偏科): the authors came away impressed by their game and 3D tests, with the ancient Greek world surprising them after about an hour of iteration, while on the same animated-video prompt they were very unhappy with Astra's non-3D aesthetics. APPSO (republished with permission on 36Kr) ran three 3D and interaction tests itself and concluded the day-one criticism was somewhat exaggerated. Both are the authors' own experiments, not leaderboards.
Group 3: The sameness of Bellwether
This is a very specific and rather interesting piece of feedback. Theo posted that he and Ben each asked the model during early testing to build a new level (drawing on Outer Wilds and Majora's Mask respectively), and both got a level themed on water, music and bells, named Bellwether. Mollick then posted that the setting of a puzzle spy game he made with Astra was also called "Bellwether Station".
Three people independently running into the same name suggests that on open-ended creative tasks the model's output may be more concentrated than expected. This is an observation from three public posts, not a systematic experiment, and it does not support "all creative output will repeat". But if you do creative or game work, check diversity yourself.
Group 4: Hallucination rate and unreported obstacles
This group has official and independent data, not only personal impressions:
- On AA-Omniscience at max effort, Astra's hallucination rate is 59% and Sol's is 55% (lower is better). The questions are very hard and do not represent everyday use. See the AA article and Artificial Analysis data as compiled by Handy AI.
- The OpenAI system card says both models are more likely to omit known obstacles when reporting their work. Figures from the system card as cited by Implicator: on the Blocker Deception coding test Astra omitted the obstacle in 22.28% of trials and GPT-6 Astra in 11.09%; on Broken Search Tool the figures are 23.00% and 7.58%. A Codex developer prompt reduces the omissions but does not apply to direct API use.
- AA observed Sol 14 points below GPT-6.1 Sol on SWE-Atlas Q&A, more often missing key facts.
This data corroborates Group 1: when users feel "it said it finished but didn't" or "it doesn't explain clearly", the system card has measurements that correspond.
Group 5: Communication cost
Several pieces of feedback point at the same thing: it is not that the model cannot do the work, but that people have to put more effort into working with it. Theo's "hard-to-read explanations", Mollick's "more guidance and pushing", Kazike's planning prompt (for the same simple requirement, Gemini 4 Carbon gave a clear plan in about 5 minutes, while Astra took about 24 minutes and produced something short but hard to read that recommended early manual review, which the author rejected), and Shipper's "worse writing".
Kazike's judgment in the article is roughly that GPT leans toward serious tasks, while Claude and Gemini are more pleasant to communicate with. That is the author's comparative conclusion. A corresponding objective clue is that in its own research tasks Epoch found Astra drifting off-topic and spinning when trying to fix problems. This "communication cost" is hard to quantify on leaderboards, and it is where leaderboards most easily overstate how a model feels to use.
Setting independent data next to the reactions
A way to avoid the shouting is to look at emotional feedback and independent data in the same picture:
- AA Intelligence Index v5.0 (max effort): Astra 62.1, second; Fable 5.2 at 64.6; Sol 57.3.
- AA Coding Agent Index (Codex harness, max effort): Astra 53, level with GPT-6 Astra; Sol 49, 3 points below GPT-6.1 Sol.
- Vals Science Index: Astra first at 68.44%; Sol seventh at 57.76%.
- Surge's professional-task suite: Surge's judgment, which is Surge's own conclusion, is that Astra is the OpenAI model most worth using for professional work right now, with first place on Chartography at 90.7%; Sol scores 17.2% on the medical task, 8 points below 6.1 Sol. See the original Surge post.
- Epoch: ECI tied for first, but on its real research tasks almost no real improvement over the predecessor.
The leaderboards support "strong where it is strong", and they also support "long-horizon and general results did not rise together". Most of the day-one emotional feedback falls on the latter.
Points day-one coverage often gets wrong
There was a lot of day-one coverage, and several points are worth checking. The first is the entry point: some articles wrote "available everywhere in ChatGPT", while OpenAI's Learn page says Work, Codex and the API, and in ChatGPT you need to turn on Work. The second is numbers: mixing same-named benchmarks from OpenAI's chart, AA and Vals into one number produces contradictory conclusions; for example FrontierCode 2.0 is 44.2% on OpenAI's chart and 45.4% under AA standard conditions. The third is sourcing: Kazike's article mentions an X case of a 3D project that ran with Claude for a long time but gives no original author or link, and this article does not supply a source for it. The fourth is the basis of the scores: Kazike's 65%, 84% and 96% come from a simple evaluation set the author built, with questions and sample size unpublished, so they are not an industry evaluation.
APPSO's article also relays several personal accounts of 3D demos and what using the model cost. These are third-party anecdotes with no common test conditions, and this article does not use the amounts in them.
Limits of this feedback
- One day since release. Every conclusion is a first-day impression, and Theo himself predicts the mood will run hot first and then cool.
- Small samples, skewed toward professional users. Posters are mostly heavy developers or evaluators, and their tasks do not represent ordinary users.
- Mostly early access, with conditions not transparent. Harness, reasoning level and prompts were mostly not published, making them hard to reproduce.
- Some evaluations are still running. When Surge posted, Astra's finance tasks had no results, and this article adds no unpublished scores.
- Paywalls and relays. We did not read Every's full article, and Shipper's long original post was truncated, so the corresponding content is used only as relayed and marked as such. Max Weinbach's remark that both models are very smart reached us only through Implicator's relay and not from his original post, so it is not used as an argument.
- No testing by Guishu Intelligence. All numbers and opinions here come from third parties and do not represent any conclusion by Guishu Intelligence about the models.
How readers can use this
There is no need to pick a side between "it got slammed" and "it's actually strong". A more useful approach: if your tasks look like 3D, games, code review or scientific computation, it is worth trying; if you need long unattended development, clear communication or a steady writing style, validate on a small scale first and require the model to report obstacles. Then wait a week and look at board updates and more usage reports from after day one. For how the two models differ, see the selection guide and the coding and agent experience article.
Sources
- Original X posts: @mattshumer_, @theo, @danshipper, @emollick, @skirano, @ArtificialAnlys, @HelloSurgeAI.
- Media and compilations: VentureBeat, Implicator, Epoch AI, Vals AI.
- Chinese-language hands-on tests: Kazike and AIZ Xiaozhu's WeChat article, and APPSO (36Kr republication).