Released on November 17, GPT-6.2 Astra and Sol delivered two conflicting report cards on their first day: third-party benchmarks put Astra at No. 2 worldwide, while early testers widely felt it was "not that impressive." The cheap part is real; whether it is good is a separate question — and that is the most interesting thing about this OpenAI release.
What the benchmarks say: Astra second, Sol does more with less
On Artificial Analysis's Intelligence Index, GPT-6.2 Astra scores 62, five points above September's GPT-6 Astra, behind Claude Fable 5.2 at 65 and ahead of Gemini 4 Carbon — good for second place; Sol scores 57.3. A Latent Space roundup highlighted the value angle: Sol matches September's Astra at roughly 15% of the per-task cost. On price, Astra is $8 per million input tokens and $40 per million output tokens, 20% cheaper than its predecessor, while Sol stays at $2 and $10; both have around 1 million tokens of context with a 128K output cap. Perplexity and OpenRouter added both models on November 17.
Why testers are unconvinced: the videos were shot last week, the models were changed in recent days
The other side of the split comes from hands-on testing. In-depth reviews say Astra's money-saving ability is real, but it shows no clear progress in coding and its communication and readability remain weak; on social media, the upbeat tone of influencer review videos contrasts with the cool reception from everyday users. Tibo, a member of the Codex team, explained that some reviewers got access early, while the team made extensive changes over the past three to four days — the videos were made last week and most likely were not retested afterward. He also acknowledged that both models are clearly better than their predecessors, but still imperfect, and the team will improve the experience in future iterations.
For users, the takeaway splits in two: on cost alone, Sol is the everyday model most worth trying right now, and Astra is the cheaper alternative for flagship-grade work; but if coding quality and readable expression matter most, second on the benchmark does not mean second in experience — it may be worth waiting for the next iteration, or testing the models on your own tasks first. OpenAI published system cards for both models on November 17, and where the next round of reviews lands will depend on whether testers are still getting the same models.