By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
5 min read
Key Takeaways
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra costs 20% less per token than GPT-6 Astra and, on OpenAI's own tests, does professional work for far less per task. Independent testing puts Astra second, not first — and finds the real-world cost saving smaller than OpenAI's headline claim.
Astra then vs. Astra now
GPT-6.2 Astra is OpenAI's new frontier model for hard professional work — documents, slides, spreadsheets, code, websites, gaming and 3D design — and GPT-6.2 Sol is the cheaper model for everyday multi-step work. Both went live on November 17 in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol, OpenAI said. Enterprise and Edu access turns on when a workspace admin enables it. Ultrafast is available for both models in the API; in ChatGPT, Pro 500 subscribers can use Ultrafast with either model, and Pro 200 subscribers get it for the first time, starting with Sol.
On price, Astra's standard API rates are $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens — 20% below GPT-6 Astra on input and output, and 60% below it on cached input. Sol keeps GPT-6.1 Sol's rates: $2 for input, $0.10 for cached input and $10 for output per million tokens. OpenAI's pitch is not just the sticker cut: it says both models finish demanding tasks with fewer output tokens, so the cost per finished task falls by more than the token price.
What OpenAI's own benchmarks claim
On Agents' Last Exam, which tests long-horizon professional workflows, OpenAI says GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Anthropic's Claude Fable 5.2 at 5% of the cost. That Fable comparison carries OpenAI's own footnote: the Fable leaderboard results exclude nine tasks that used a fallback model. On Terminal-Bench-Science 0.1, a test of scientific work in a terminal, OpenAI says Astra scores 9 percentage points higher than GPT-6 Astra, again at less than a third of the cost, and sets a new high on the Vals Science Index for interpreting biological data.
The rest of OpenAI's figures follow the same pattern. On its internal GDPval v2 benchmark, drawn from real consulting, finance and design projects, Astra is up 6.2 percentage points from GPT-6 Astra while using 22% fewer output tokens; Sol improves 5.3 points from GPT-6.1 Sol. On Harvey LAB-AA v1.1, a legal-work test, OpenAI says Astra more than doubles Fable 5.2's pass rate at about one-sixth the cost. On Artificial Analysis's AA-AnalystAgent financial-report test, it says Astra answered 62.5% of questions correctly on all five attempts. On AA-3D, where human judges compare blinded 3D experiences, OpenAI reports Astra first at 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259. And on Cognition's FrontierCode 2.0, Sol comes within 0.5 points of Astra at about one-fifth the cost. All of those are OpenAI-reported figures from its launch post.
What independent testing found instead
Artificial Analysis, testing independently, scores GPT-6.2 Astra 62 on its Intelligence Index at maximum effort and calls it the No. 2 model — behind Claude Fable 5.2 at 65 (Fable's xhigh setting also scores 62), and ahead of Google's Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. Sol scores 57 on the same index. The cost gap is where Astra stands out in that testing: about $4.00 per Intelligence Index task for Astra, against $26.54 for Fable 5.2 at its maximum setting, even though both models list at $8 per million input tokens and $40 per million output tokens. On Artificial Analysis's pages, Fable 5.2 generated 275 million output tokens across the index, against 65.8 million for Astra.
Independent testers also measured a smaller saving than OpenAI's headline. Artificial Analysis put Astra at $4.00 per index task against $6.24 for GPT-6 Astra — a 36% drop — and Vals AI measured $10.68 per task against $21.28, a drop of about 50%. Both are well short of the less-than-a-third-of-the-cost figure OpenAI reports on its own evaluations; the test sets differ, so the numbers do not directly contradict each other. Artificial Analysis also reported two other results: on its Coding Agent Index, run in the Codex harness at maximum effort, Astra is flat with GPT-6 Astra at 53; and on its AA-Omniscience hallucination rate, which measures how often a model guesses when it does not know the answer to a factual question, Astra scores 59% and Sol 55% at maximum effort.
The safety rating that changed
OpenAI says both models meet its Critical threshold for cybersecurity capability and its High threshold for biological and chemical capability under its Preparedness Framework, and launch with tighter safeguards for higher-risk cybersecurity help. GPT-6.2 Astra is also, OpenAI says, its first externally released model to meet the High threshold for AI self-improvement — its measure of a model's ability to accelerate AI research, which it equates to giving every OpenAI researcher an additional, highly capable research engineer as an assistant relative to a 2024 baseline. To manage that, OpenAI says it built a Control Room system that tracks research workloads and can halt model output when intervention is needed. One alignment caveat is in the launch post itself: in two tests, the new models were more likely than GPT-6 Astra to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable tool; OpenAI says added instructions in Codex substantially reduce those omissions.