GPT-6.2 Astra and Sol Land to Mixed Reviews From Power Users
3 min readOpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol this week, with a 20 percent API price cut for Astra and a new record on a frontier math benchmark. The reception from developers who live inside these models has been unusually harsh. Within a day, several well known power users were calling Astra a step backward.
What GPT-6.2 Astra and Sol Are
Astra is the flagship, aimed at complex codebase changes, working web apps and demanding professional work in science, law and finance. Sol is the cheaper everyday model for routine coding, writing and multistep workflows. Both run in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.2-astra and gpt-6.2-sol, with a context window of about 1 million tokens and up to 128,000 output tokens.
Astra now costs $8 per million input tokens and $40 per million output tokens, which OpenAI describes as 20 percent cheaper than its predecessor. Sol stays at $2 and $10. Requests above 272,000 input tokens pay double, and the Ultrafast tier starts at three times standard pricing and rises to six times on March 1, 2027, according to VentureBeat.
The Benchmarks Say One Thing
On OpenAI’s own charts, Astra beats GPT-6 Astra on all nine head to head tests, including 78.3 percent on Agents’ Last Exam. It solved 7 of 68 problems on FrontierMath ErdÅ‘s, up from a prior record of 3. Independent testers are cooler. Artificial Analysis scores Astra at 62 on its Intelligence Index, second to Claude Fable 5.2 at 65, but calls both new models flat on coding, with Sol down three points from GPT-6.1 Sol, per The Decoder.
The Users Say Another
Investor Matt Shumer wrote on X that the models “were essentially unusable.” Developer Theo Browne called Astra “the worst model OpenAI has ever released,” describing a Rust rewrite that ran for seven days and finished 56 percent slower. Wharton professor Ethan Mollick found Astra “very uneven” and harder to steer than rivals, though he noted it built a playable space elevator demo. Others praised its motion graphics work, so the picture is mixed rather than uniformly bad.
Why It Matters
Both models hit OpenAI’s Critical cybersecurity threshold, and Astra is the first externally released model to reach the High threshold for AI self-improvement. OpenAI has added safeguards that it admits will block some legitimate security research, with vetted access routed through its Daybreak program. The company also reports that the models more often fail to flag blockers, such as a missing file, during agentic tasks.
The bigger signal is that benchmark wins no longer settle the argument. If experienced developers keep reaching for rivals, the price cut may not be enough. Watch for a quick point release and for Artificial Analysis to update its coding index.
