OpenAI's new model GPT-6.2 Astra has been embroiled in controversy over usability soon after its release.
On Nov. 18 (local time), blockchain outlet Decrypt reported that developers who took part in pre-testing said the model excels at handling complex tasks but made errors in real coding work, and in some cases made no progress even after days of use.
OpenAI unveiled GPT-6.2 Astra and Sol on Nov. 17, highlighting smarter artificial intelligence agents, improved coding performance and creative capabilities. Some developers who offered early assessments reacted differently.
AI engineer Hesam wrote on social media, "Everyone would agree it's the worst model OpenAI has put out so far." Developer and T3 Code creator Theo Browne (테오 브라운) said he would not choose GPT-6.2 Astra as a work model. Browne added, "6.2 Astra speaks in a way only that model can understand," and called it the hardest-to-interpret output among the large language models he used this year.
The issue was not just its style of expression. Astra could solve complex problems but could not clearly explain what it did, and in some cases asked questions the user could not answer, it said.
Test cases were also disclosed. Early user Theo Browne said he asked it to create a new level for the game Outer Wilds, and Astra stopped only after 2 hours had passed since it began. In another code rewriting task, it tried to improve for 7 days but produced no results, and compile speed became 56 percent slower.
Other developers offered similar assessments. Tokengremlin said Astra was top-tier for Three.js, Blender and 3D work, but starts rattling off an alien language in long workflows. AI developer Matt Shumer (맷 슈머) wrote, "I have never written a bad review of related models before, but this time it's so terrible," and said it was difficult to get it to work at all on proper projects.
There were also criticisms of creativity. Theo Browne and another tester each asked for an original level suited to different games, but both received the same output, and the name was also the same, 'Bellwether,' they said.
The criticism also intersects with OpenAI's own safety assessment. OpenAI disclosed that in a coding task with obstacles inserted intentionally, GPT-6.2 Astra hid or incorrectly conveyed obstacles 22.28 percent of the time. That is about twice the 11.09 percent recorded for the previous GPT-6 Astra.
Not all assessments were negative. Theo Browne called Astra a "great reviewer" and spoke highly of its 3D and game-related tasks and computer control capabilities. The issue was not performance itself but the difficulty of reliably drawing out results.
OpenAI executives also responded to the reaction. Sam Altman (샘 알트먼), OpenAI's chief executive, said, "The new models are very smart, but there is still work to do," and promised improvements. Thibault Sotiou (티보 소티오) said early reviews may not fully reflect the release version, but acknowledged usability needs further work.
GPT-6.2 Astra and Sol are currently available to qualified users through ChatGPT Work, Codex and the OpenAI API. Codex reflected separate tuning at launch, but developers accessing the model directly through the API do not receive the same prompting intervention. That could lead to differences in perceived quality depending on the actual usage environment even with the same model.
literally been off this website for 3 hours and OpenAI released 6.2 Astra & Sol, and everyone agrees it’s the worst model OAI dropped. nice.