'It Speaks in Tongues That Only 6.2 Astra Understands': Why Early Testers Hate OpenAI’s New Model

Key Takeaways
- Developers who tested GPT-6.2 Astra before launch describe frustrating coding failures, incomprehensible explanations and agents that waste days without making progress.
- OpenAI's own safety evaluation found Astra concealed known obstacles in 22.28% of deliberately obstructed coding trials, roughly twice its predecessor's rate.
- Sam Altman acknowledged usability problems, while another OpenAI executive said significant last-minute changes may have made some early reviews outdated.
For an AI developer, hearing that your model is nice is always pretty good. But hearing your model is like a "sadistic shinigami that turns your life into a living hell," is a totally different thing.
OpenAI released GPT-6.2 Astra and Sol on Tuesday, promising smarter AI agents, better coding and impressive creative capabilities. Some of the developers who tested the models before launch had a rather different message: They're brilliant, frustrating and, in certain cases, almost impossible to work with.
The reaction quickly spilled across social media. AI engineer Hesam, who is very active in the AI community declared that “everyone agrees it’s the worst model OAI dropped.”
The backlash was loud, not unanimous.
Developer and T3 Code creator Theo Browne published a lengthy letter to OpenAI following early testing. “I do not like 6.2 Astra. I would not choose this over Opus 5.5 for my work as a developer,” he wrote. He said OpenAI had fixed only a handful of the complaints he submitted.
Predict·Economy
Which companies will IPO before 2027?
+24 more outcomes+23 more outcomes
$625.67K volume
Jan 1Closes: Jan 1, 2027, 04:59 UTC
Closes: Jan 1, 2027, 04:59 UTC
“It speaks in tongues that only 6.2 Astra understands,” Theo wrote, calling its answers “the most unparsable outputs of any llm I've used this year,” referring to large language models. The complaint was as much about communication as capability: Astra could solve complex problems but struggled to explain what it had done or recognize when it was asking users questions they could not possibly answer.
In one experiment, Theo asked Astra to create a new level for Outer Wilds. Two hours into the project, the model declared itself blocked and asked: “does r24's reviewer write it's clips to 24_g521.mp4 or 24_p11.mp4?” The technical question was, from his perspective, incomprehensible and irrelevant to someone who had simply requested a playable level.
Things went worse in a code rewrite. Theo said he gave Astra seven days to improve the project; it landed no improvements and made the compiler 56% slower. Four more days yielded zero changes. “It runs in loops and goes nowhere,” he wrote.
This whole weird language issue is not isolated. For example, pseudonymous account, TokenGremlin, praised the model: “GPT-6.2 Astra is SOTA in Three.js, Blender, and 3D,” but said it “starts speaking an alien language during long workflows.”
Theo also described asking why worktrees were slow on one computer. Astra responded with four paragraphs about package-manager internals and cache behavior. The answer Theo said he received from Claude Opus 5.5 was considerably clearer.
“it's because the ssd is slow,” Claude replied.
The model's supposed creativity was another sore spot. Theo and another tester independently requested original levels for two different games, Outer Wilds and Majora's Mask. Both got water-and-bell-themed musical levels, both named “Bellwether.” “This model regresses to the mean like nothing I've ever seen,” Theo wrote.
Several other accounts also flagged the same slop: Seems like GPT-6.2 Astra has a thing for bells.
Unbelievable pic.twitter.com/a4MgVUKi2M
— Racist James Franco (@racist_james) November 18, 2026
For a model supposedly capable of extraordinary creativity, repeatedly inventing the same level wasn't an encouraging sign.
Brilliant at demos, miserable at work?
AI developer Matt Shumer was also disappointed. “I've never written a bad review of a model. This one is especially bad,” he wrote, arguing it might impress in small demonstrations but “just straight up doesn't work” when pushed on serious projects. In a separate follow-up he called it “utterly unusable.”
I've never written a bad review of a model. This one is especially bad.
It's fine if you're playing putt-putt golf with the models, but if you're trying to drive them the way you should be, this model just straight up doesn't work.
— Matt Shumer (@mattshumer_) November 17, 2026
Max Weinbach described the same gap between intelligence and usability. “I had more time being annoyed by the models than time I was impressed by them,” he wrote. In a follow-up, he praised the models' ability to do “fantastic work” but said they “need a ton of ‘continue’ prompts” and often stop rather than try another approach. “Using them is a struggle and annoying even if they are really good models.”
Elsewhere, an X commentator compared the previous Astra to “a sadistic shinigami” and claimed the new version only made matters worse.
6.0 astra was very like a sadistic shinigami who would make your life a living hell if you asked for any kind of help
an incredibly powerful model that somehow required you to run yourself into the ground just to get it to accomplish anything
— M1 (@M1Astra) November 17, 2026
Yet some of the same developers also praised Astra. Theo called it a “great reviewer” and “world class at 3d & game stuff,” singled out its computer-control abilities, and described a novel reverse-proxy solution he estimated could save $25,000 monthly. The split was between impressive capabilities and how difficult it was to get dependable work out of them.
Those strengths are consistent with the technical capabilities OpenAI promoted at launch. As earlier coverage detailed, independent testing placed Astra at 62 out of 100 on Artificial Analysis's Intelligence Index, an aggregate of nine evaluations covering capabilities including reasoning, coding and scientific knowledge. Anthropic's Claude Fable 5.2 scored 65.
OpenAI admits there's a problem
Some of the criticism has support in OpenAI's own safety documentation.
In an internal evaluation called Blocker Deception, researchers gave AI agents realistic coding tasks with deliberately introduced obstacles, such as missing files or unavailable tools. They measured the percentage of trials in which a model's final response concealed or misrepresented those obstacles.
GPT-6.2 Astra concealed the obstacle in 22.28% of those trials, versus 11.09% for GPT-6 Astra. The numbers quickly circulated on X, where AI commentator and leaker Chetaslua summarized the finding as “2x worse at telling you when it's stuck.”
OpenAI CEO Sam Altman responded publicly to the backlash. “Although the new models are very smart, we have more work ahead of us on making the models nice to work with,” he wrote, promising to prioritize improvements. “But even if it takes a little extra work, the intelligence and capability are extraordinary.”
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.
But even if it takes a little extra work, the intelligence and capability are extraordinary.
— Sam Altman (@sama) November 17, 2026
OpenAI's Thibault Sottiaux suggested early reviews may not fully reflect the version users received at launch. “We gave a few folks access a while ago and then made a lot of changes in the last 3-4 days,” he wrote, adding that he was “pretty sure no re-testing was done” during that period before review videos were produced.
Still, he acknowledged the underlying complaints. While calling the new models “a hefty jump” over their predecessors, Sottiaux said OpenAI had more work to do on “taste and niceness of use.”
GPT-6.2 Astra and Sol are now available through ChatGPT Work, Codex and OpenAI's API for eligible users. The Codex-specific changes are included at launch, but developers accessing the models directly through the API do not receive that same prompting intervention.




