Models

OpenAI's GPT-6.2 Astra and Sol draw backlash from early testers

OpenAI's GPT-6.2 Astra and Sol beat predecessor benchmarks, but Wharton's Ethan Mollick calls Astra "harder to steer" and "lazy" — while OpenAI's Codex lead says testers may have reviewed outdated builds.

OpenAI’s GPT-6.2 and Sol Are Getting Backlash From Early Testers
OpenAI’s GPT-6.2 and Sol Are Getting Backlash From Early TestersPexels / pixabay
By Sophie Lindqvist5 min read

Why it matters

  • OpenAI released GPT-6.2 Astra and GPT-6.2 Sol this week, both surpassing predecessor benchmarks, available to paid subscribers via ChatGPT Work and Codex.
  • Wharton's Ethan Mollick said GPT-6.2 Astra was "harder to steer and work with" and "frequently seemed 'lazy'" compared with competitor models.
  • NETSYNC head of AI Jason J. Fleagle called GPT-6.2 "one of the worst AI models OpenAI has ever released."
  • OpenAI Codex leader Thibault Sottiaux said early builds went to influencers earlier and models were updated days before release, with no retesting assumed.
  • OpenAI did not immediately reply to a request for comment.

OpenAI released two new flagship models this week — GPT-6.2 Astra and GPT-6.2 Sol — and both beat their predecessors on benchmarks, yet early testers are already pushing back hard on how the models behave in real work.

The criticism isn't coming from casual users. It's coming from developers, AI researchers, and consultants who stress-test models professionally — the audience OpenAI arguably cares about most. That gap between benchmark gains and day-to-day usability is the story here, and it's a pattern that has haunted every major model release this year.

What did OpenAI actually ship?

The release follows OpenAI's established two-tier structure. Astra is the flagship built for complex tasks, the model you reach for when a problem needs depth. Sol is positioned as the "everyday" model — lighter, faster, aimed at routine work. Both are currently limited to paid subscribers and available through ChatGPT Work and Codex, OpenAI's coding-focused product.

On paper, both models surpass the benchmarks of their predecessors, according to the company's release. For the average person firing off questions in ChatGPT, the 6.2 Sol and Astra pair should perform more than serviceably, as CNET reports. The complaints start where the professional work starts — users who dig into the models, run structured tests, and build on top of them are reporting hiccups they're not used to seeing from OpenAI.

The takeaway: the models shipped, the benchmarks improved, and the rollout is paid-users-only for now.

Who is criticizing the models, and what did they say?

The most damaging assessment came from Ethan Mollick, associate professor and director of the Generative AI Lab at the University of Pennsylvania's Wharton School of Business. In a LinkedIn post, Mollick wrote that GPT-6.2 Astra was "harder to steer and work with," and "frequently seemed 'lazy'," compared with competitor models.

That word — "lazy" — matters. When testers call a model lazy, they usually mean it cuts corners on complex tasks, stops short of full answers, or requires more prompting to get the same output quality. Coming from Mollick, who runs one of the most closely watched academic labs studying generative AI in practice, the criticism carries weight well beyond a typical user complaint.

The second quote is blunter. Jason J. Fleagle, head of AI at NETSYNC, wrote on LinkedIn: "I think GPT-6.2 has to be one of the worst AI models OpenAI has ever released."

Two data points don't make a consensus. But both are named, senior professionals posting under their own names, and both point in the same direction: a model that scores well on benchmarks but frustrates the people trying to use it for demanding work.

The takeaway: credible, named testers — including a Wharton lab director — say the new models regress on usability, not raw capability.

Is there an explanation for the gap?

There may be, and it involves the mechanics of how OpenAI seeded these releases. Thibault Sottiaux, known as Tibo, leads OpenAI Codex, and he addressed the backlash in a post on X.

According to Sottiaux, early builds of 6.2 Sol and Astra were released to influencers some time ago. Then, within the past couple of days, OpenAI made updates and changes to the models. His assumption, as he stated it, is that no retesting was done since those changes landed.

If that's accurate, it explains a lot. Influencers and early testers may have formed their judgments on older builds, published those judgments, and never circled back to re-evaluate the updated versions. Meanwhile, the negative takes spread while the models themselves kept changing underneath. It's a version of the classic evaluation-lag problem in AI releases: the review cycle can't keep pace with silent model updates.

Sottiaux didn't claim the updates definitively fix the problems. As CNET notes, only time and more testing will show whether these improvements make much of a meaningful difference.

The takeaway: OpenAI's own Codex lead says testers may be reviewing stale builds — which muddies every verdict published so far.

Why this matters beyond OpenAI

A new AI model seems to drop almost every couple of days from one of the big tech giants, each improving on the previous version in one way or another. That release cadence is exactly what makes stories like this one recur. When models ship faster than anyone can properly evaluate them — and then get silently updated after launch — the whole ecosystem of reviews, benchmarks, and word-of-mouth gets shakier.

For OpenAI specifically, the stakes are competitive. Mollick's comparison to "competitor models" signals that professional users now evaluate each release against rivals, not just against OpenAI's own previous generation. A model that's harder to steer than the alternative loses that comparison even if its benchmark numbers win.

OpenAI didn't immediately reply to a request for comment from CNET on the backlash.

(Disclosure note from the original reporting: Ziff Davis, CNET's parent company, in 2025 filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

What happens next

The open question is whether the post-release updates Sottiaux described actually close the gap between the benchmarks and the user experience. Expect a second wave of testing on the updated builds in the coming days — and expect OpenAI's response, or silence, to shape how GPT-6.2 is remembered.

Source: cnet.com

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

823 articles

Related articles

  1. OpenAI's GPT-6.1 Sol Nears Flagship Astra at One-Fifth the Price
  2. OpenAI Ships GPT-6.1 Sol at a Fifth of Astra's Price
  3. OpenAI’s GPT-6.2 Astra earns a High self-improvement rating
  4. OpenAI launches GPT-6.2 Astra and Sol, cuts API prices 20%
  5. OpenAI Launches GPT-6 Sol and Luna at Half the Price

« Previous articleNext article »