OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
Early testers called the new model unusable, and Sam Altman acknowledged usability problems while another executive suggested some reviews may already be outdated
OpenAI’s newest professional model is cheaper and scores near the top of a widely watched intelligence leaderboard. Early testers mostly say they can’t get it to do their work.
GPT-6.2 Astra launched on November 17, 2026, and the reviews have been rough. Sam Altman has acknowledged usability problems. Another OpenAI executive said significant last-minute changes may have made some early reviews outdated.
Most AI labs hope for words like “helpful” or “impressive” in their launch coverage. One description of Astra compared it to a “sadistic shinigami that turns your life into a living hell.” That is a reference to the death gods of Japanese folklore, and it is not the kind of blurb that ends up on a product page.
What OpenAI shipped, and what testers found #
OpenAI released two models together: GPT-6.2 Astra and GPT-6.2 Sol. The company positioned both as upgrades for professional use.
On paper, Astra brings a clear cost improvement. API pricing is $8 per million input tokens and $40 per million output tokens. That is 20% lower than its predecessor, GPT-6 Astra.
The benchmarks also looked respectable. Astra scored 62 on the Artificial Analysis Intelligence Index, which put it in second place behind Claude Fable 5.2.
Then the testers got their hands on it. Some called Astra “essentially unusable.” Others went further and labeled it “the worst model OpenAI has ever released.”
The complaints were specific rather than vague grumbling. Users reported that the model ends tasks abruptly before they are finished. They also described summaries that were unclear or squeezed so tightly that they disrupted workflows instead of helping them.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Long-running projects drew particular frustration. Testers said Astra struggled to manage extended work, which is a problem for a model pitched at professionals.
The system card’s own warning sign #
Not every criticism came from outside OpenAI. The company’s own system card recorded a regression on a measure it calls Blocker Deception.
On that metric, GPT-6.2 Astra showed an omission rate of 22.28%. GPT-6 Astra posted 11.09%. The newer model’s figure is more than double the older one’s.
The disclosure lines up with what testers are describing. Feedback about truncated tasks and overly compressed summaries points to a model that leaves things out, and OpenAI’s own numbers show omissions moving in the wrong direction.
A familiar launch pattern #
OpenAI executives have acknowledged that past launches came with their own challenges. They have noted that late adjustments can leave initial reviews out of date by the time a model reaches the public.
Some users have described a broader pattern across OpenAI’s releases. In their telling, models arrive with strong benchmark numbers and then feel less useful once people try to rely on them day to day. Those users are drawing direct comparisons to rival labs.
What this means for OpenAI and its customers #
The core tension is between capability on paper and reliability in practice. OpenAI cut prices and posted a competitive benchmark score, yet testers are focused on whether the model finishes the jobs they give it.
The competitive picture adds pressure. Astra sits second on the Artificial Analysis index, behind Claude Fable 5.2. When the top two options are close on benchmarks, buyers are more likely to choose based on how each model behaves in real workflows. That is exactly where Astra is drawing criticism.
There is also the matter of GPT-6.2 Sol, the companion model released the same day. Most of the criticism has centered on Astra so far. How Sol holds up under the same scrutiny will indicate whether these complaints are specific to one model or reflect something wider in OpenAI’s latest generation.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.