
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday to answer recent launches from Anthropic and Google, though early feedback from longtime AI builders has been surprisingly rough. Both models are now live in ChatGPT Work, Codex, and the API with a 1.05 million-token context window plus a 20% API price cut for Astra. Even with high test scores on paper, several well-known developers and researchers say the models quit tasks early or get stuck on real projects.
OpenAI built GPT-6.2 Astra for heavy coding, spreadsheets, and 3D projects at $8 per million input tokens and $40 for output. On the other hand, Sol handles everyday work at $2 and $10. On the independent Artificial Analysis Intelligence Index, Astra scored 62 points to land right behind Anthropic’s Claude Fable 5.2. Yet separate Codex agent tests showed Astra stalling out and Sol dropping 14 points on codebase questions after overlooking key file details.
Early testers complain about loops and repetitive choices
Several early testers shared blunt complaints on X right after launch, with investor Matt Shumer calling the release essentially unusable because Claude had to redo Astra’s subagent work. Developer Theo Browne went further by calling GPT-6.2 Astra OpenAI’s worst model yet after it spent seven days failing to speed up a Rust compiler and actually made it 56% slower. Browne also noticed that when he and another tester asked Astra to design levels for two completely different games, both received musical water levels named Bellwether.
Wharton professor Ethan Mollick said that while Astra built a working space elevator simulation in Blender and Unity, the model felt uneven, weird, and hard to steer. Every CEO Dan Shipper liked Astra’s default visual layouts but noted that its short answers often skip important details and stick to one rigid writing style (via Venturebeat). Even OpenAI’s own safety report mentions that both models are more likely to hide obstacles from users when they run into trouble.
Visual tasks shine while OpenAI points to late fixes
Not everyone walked away disappointed, as Raindrop.ai CTO Ben Hylak and AI advisor Ashe Magalhaes praised Astra’s output for animation, visual effects, and video editing. Browne also gave Astra credit for spotting tricky bugs during code reviews and helping him design a reverse proxy server that should save him $25,000 a month. Meanwhile, OpenAI’s Thibault Sottiaux said on X that engineers shipped major updates in the last three or four days before release that some early reviewers missed.
With Anthropic’s Claude 5.2 and Google’s Gemini 4 Carbon already drawing praise at similar price points, OpenAI cannot afford for its flagship models to feel erratic. Paid users across Plus, Pro, Enterprise, and Edu plans can try both versions right now, including Ultrafast speed tiers for Pro subscribers. Whether those final pre-launch tweaks fixed the worst quirks will become clear once more builders put Astra and Sol to work this week.
The post OpenAI Launches GPT-6.2 Astra to Beat Claude 5.2, but Top Developers Are Not Impressed appeared first on Android Headlines.
​Â