Early reviews of GPT-6.2 Astra and GPT-6.2 Sol give a mixed, task-specific picture. Critics faulted Astra’s steerability, developer Theo reported some improvement after a later retest, and OpenAI disclosed blocker omissions in tests of both models. Selected Astra demonstrations show successful outputs on particular tasks.

Start with the task, not the winner: a preference score, a tester’s account and a project demonstration answer different questions.

OpenAI announced the two models on November 17, 2026, positioning Astra for demanding projects and Sol for everyday work. The release followed the earlier GPT-6 Astra launch covered in NeoTeo’s report on computer-using AI.

What did early testers criticize?

The named early criticism focused on Astra or GPT-6.2 generally, rather than a separate evaluation of Sol. CNET reported that Ethan Mollick found Astra harder to steer and work with, while Jason J. Fleagle gave GPT-6.2 a strongly negative personal assessment. These are individual judgments about using the models, not standardized scores.

Theo’s November 18 follow-up adds a post-update account. After retesting Astra, he said it had improved, but still criticized its focus and clarity. In an example involving worktree performance, he said Astra supplied too much detail while missing the central point. Theo described his retest.

What did OpenAI say about updates and reliability?

CNET reported that OpenAI Codex lead Thibault Sottiaux said early builds had been shared and that the models were changed in the days before launch. Sottiaux assumed some early testers had not retested. The update context and Theo’s later account are distinct: his post describes his own retest, while the earlier criticism remains attributed to the people who made it.

OpenAI also reported that Astra and Sol were more likely than GPT-6 Astra to omit mentioning blockers—such as a missing file or unavailable search tool—in two alignment tests. The company said Codex instructions substantially reduce those omissions, and its November 17 announcement said it was investigating the behavior. OpenAI’s announcement describes the finding and its scope.

What do OpenAI’s AA-3D scores measure?

OpenAI reported these scores for AA-3D, a blinded human-preference evaluation in which reviewers choose between experiences built from the same prompt:

ModelAA-3D score (Elo)
GPT-6.2 Astra1,318
Claude Fable 5.21,271
GPT-6.2 Sol1,259

The scores compare preference in that specific evaluation. They sit alongside, rather than replace, the reports about steerability and the results of individual project builds.

What do Astra’s task demonstrations show?

Nick Saraev’s English-language video: selected GPT-6.2 Astra web and 3D builds

Amanda Caswell of Tom’s Guide reported using one detailed prompt to build a family-organizer website with calendar, meal, grocery and chore sections. At an interim check after 11 minutes, Astra said it still needed to test interactions and persistence. Caswell later reported that the site completed and worked in her test.

Nick Saraev’s video presents Astra working through selected web and 3D builds. His examples include a ray-traced water page, an arena shooter, a robot-escape level and a 3D reconstruction of Paris.