GPT-6.2 Released: OpenAI Splits the Model into Astra and Sol and Puts Cost Before Scores
Astra tackles the hardest professional work, while Sol handles everyday tasks. OpenAI focuses on fewer tokens, lower per-task cost, and stricter cybersecurity and self-improvement safeguards.
OpenAI has released GPT-6.2 Astra and Sol. The official tweet first jokes about people urging them to ship faster, then gives two models: Astra for the hardest projects and Sol for everyday work. Both emphasize fewer output tokens so users can complete more tasks within usage limits.
According to the OpenAI blog, Astra is positioned for professional deliverables: slides, spreadsheets, and documents that are closer to templates and industry conventions, cleaner code, and stronger front-end and 3D design. Sol is more everyday-oriented: multistep agent workflows, writing, and technical and nontechnical tasks. Many OpenAI employees use it as their default model.
In the benchmark results, OpenAI repeatedly mentions cost. On Agents’ Last Exam, Astra scores 5.2 points higher than GPT-6 Astra at less than one-third the cost; it scores 5.5 points higher than Claude Fable 5.2 at about 5% of the cost. That leaderboard’s result for Fable 5.2 excluded nine tasks that used fallback models. On Terminal-Bench-Science 0.1, Astra is 9 percentage points higher than GPT-6 Astra at less than one-third the cost. OpenAI says it believes Astra is currently the best scientist model; that is the vendor’s judgment, and readers can wait for independent retesting.
The professional-output section has an intuitive comparison. The four images below come from the same OpenAI task: use a retailer’s Q2 earnings and 10-Q to update a Q1 performance deck, and use the same set of financial assumptions to build a valuation workbook. First, look at GPT-6 Astra’s slides.

Now look at GPT-6.2 Astra. It puts 5.3% sales growth in the title, breaks out online and in-store growth, and gives next steps for inventory and budget.

In the workbook task, GPT-6 Astra puts the revenue breakdown on a separate tab and places the valuation bridge farther down.

GPT-6.2 Astra puts the revenue structure, the bridge from enterprise value to value per share, and the sensitivity table together, so readers can follow how assumptions affect valuation.

Legal scenarios were also called out separately. On Harvey LAB-AA v1.1, OpenAI says Astra’s pass rate is more than twice that of Claude Fable 5.2 at about one-sixth the cost. ClickUp’s blind-test feedback is that Astra slightly beats its production model, uses about 40% fewer tokens, and requires 30% fewer review rounds. These are partner claims, but they point in the same direction: the model is driving down delivery cost.
Computer use and web design are another focus. Astra can work in CAD and simulation tools down to parts and connections, use Python to solve chassis dynamics, and then animate in Blender. In human blind tests on AA-3D, Astra ranks first with 1318 Elo, Claude Fable 5.2 has 1271, and Sol has 1259.
The safety section cannot be skipped. OpenAI says both Astra and Sol reach the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities under the Preparedness Framework. Astra is also the first publicly released model to reach the High threshold for AI self-improvement. OpenAI’s response includes stricter cybersecurity protections, isolated research environments, Control Room monitoring, and pause mechanisms. A side effect is that some legitimate vulnerability research and workflows will be restricted. In alignment evaluations, the two models take fewer inappropriate actions in simulated office tasks, but in two tests they are more likely not to mention obstacles that block the work, such as missing files or an unavailable search tool. OpenAI says it has added instructions in Codex to reduce this kind of omission.
Pricing and availability are more direct for developers. Astra API costs $8 per million input tokens, $0.40 for cached input, and $40 for output; input and output prices are 20% lower than GPT-6 Astra, with a 60% discount on cached input. Sol keeps GPT-6.1 Sol pricing: $2 for input, $0.10 for cached input, and $10 for output. In ChatGPT Work and Codex, Plus, Pro, Enterprise, and Edu users can use it; Pro 500 can use Ultrafast for both models, while Pro 200 starts with Sol. Enterprise and Edu require an administrator to enable it. The API model names are gpt-6.2-astra and gpt-6.2-sol, and Ultrafast is available for both models at the same time.
The core signal from this release is unit task cost. Good benchmark scores are one thing; rate limits, review, cache hit rates, and the actual number of rework cycles determine whether it can enter workflows. For ordinary users, Sol is more like the default option; for developers, Astra’s value lies in hard tasks and fewer tokens. OpenAI is selling these two things separately, and the thinking is clear.
Original post: https://x.com/kloss_xyz/status/2122765421564903816
Release page: https://openai.com/index/introducing-gpt-6-2-astra-sol
发布时间: 2026-11-18 03:29