OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, rolling them out to ChatGPT Work and Codex users on Plus, Pro, Enterprise and Edu plans, and to developers through the API. Astra is priced at $8 per million input tokens, $0.40 for cached input and $40 for output. Sol keeps its predecessor's pricing of $2, $0.10 and $10. Both models offer a 1 million token context window and a 128,000 token output limit. On the Artificial Analysis Intelligence Index, Astra scores 62, just behind Anthropic's Fable 5.2 at 65. Artificial Analysis also found that Sol matches the earlier September Astra model at about 15 per cent of the cost per task.
OpenAI's own benchmark claims are broad. It says Astra, run at its highest reasoning setting, beat GPT-6 Astra by 5.2 points on Agents' Last Exam 1.1 while using less than a third of the cost per task, and that it used 22 per cent fewer output tokens on GDPval v2. Third-party results are more mixed. Mercor's OpenWorldEngineer test put Astra at 67.2 per cent, against 54.6 per cent for Fable 5.2, and Artificial Analysis ranked Astra first on its 3D benchmark. OpenAI also rated Astra 'High' for AI self-improvement under its preparedness assessment, the first externally released model it has been given that rating.
Early reactions from testers are sharply divided. Developer Matt Shumer called the models essentially unusable after weeks of testing, saying agents built on Anthropic's Opus and Fable models rejected much of Astra's subagent work. Pietro Schirano attributed many of his early problems to the surrounding harness rather than the model itself. The Latent Space newsletter, which reported the launch, said some of the criticism looked exaggerated but that unusual reinforcement-learning artefacts in Astra were noticeable. For Australian businesses and professionals, cost is the main practical question. Astra costs four times as much as Sol on both input and output, so teams running high volumes should test whether Sol is good enough before defaulting to Astra. Any workflow that sends customer or staff personal information to either model still falls under the Privacy Act 1988, including its security and disclosure obligations, so check how data is handled under the API terms. Staff who only use ChatGPT on existing plans will see the new models, but their obligations do not change.
Key Points
Astra scores 62 on the Intelligence Index, behind Fable 5.2 at 65, while Sol matches the earlier Astra at about 15% of the cost per task.
Astra costs $8 input and $40 output per million tokens, four times Sol's $2 and $10, and both models offer a 1M token context window.
OpenAI rated Astra 'High' for AI self-improvement, a first for one of its externally released models.
Early tester reactions are sharply split, with some calling Astra essentially unusable and others blaming the harness.
Vendor-reported benchmark gains on long engineering runs and 3D tasks still need independent confirmation.