GPT-6.2 Released: Astra Brings Costs Down and Raises the Cybersecurity Bar
OpenAI releases GPT-6.2 Astra and Sol. One tackles the hardest tasks, the other takes over the daily workstation; the company emphasizes fewer tokens and lower costs while tightening high-risk cybersecurity requests.
On November 17, 2026, OpenAI released GPT-6.2 Astra and Sol. Astra targets the hardest projects, while Sol targets everyday work. The official tweet begins: “Some of you have been asking very politely. Some of you have just been asking.” Some people online said the line hit their funny bone, while others commented that this may be OpenAI’s coldest launch post. The copy is indeed short this time, with the focus on efficiency and cost.
The official core promise: both models get better results on demanding tasks with fewer output tokens, and the same usage limits can accomplish more. For enterprises and developers, this means the cost of frontier performance can be sustained at scale.
Astra’s benchmarks: on Agents' Last Exam, it 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 5% of the cost. OpenAI calls it the best model for scientists: on Terminal-Bench-Science 0.1 it is 9 percentage points higher than GPT-6 Astra at less than one-third the cost; it also sets a new high on the Vals Science Index. These scores come from the company; details are in the system card and benchmark footnotes.
Sol is positioned more practically. OpenAI says it is smarter and more efficient than GPT-6.1 Sol, can run multi-step agentic workflows, and writes naturally. Many OpenAI employees use it as their default model. On FrontierCode 2.0, Sol is 0.5 points behind Astra at about one-fifth the cost; on LifeVal it surpasses GPT-6 Astra at about one-seventh the cost.
Pricing: 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, and cached input is 60% lower. Sol continues GPT-6.1 Sol: $2 input, $0.10 cached input, and $10 output. OpenAI recommends keeping the prompt prefix unchanged and appending new instructions to improve cache reuse and reduce cost and latency. This point is more useful to developers than the promotional images.
Availability: Starting on launch day, Plus, Pro, Enterprise, and Edu users can use it in ChatGPT Work and Codex. Pro 500 can enable Ultrafast for both models; Pro 200 gets Ultrafast for the first time, starting with Sol. The API provides access to gpt-6.2-astra and gpt-6.2-sol, with Ultrafast available at launch.
Professional deliverables: Astra adheres more closely to industry templates in slides, spreadsheets, and documents. On legal tasks in Harvey LAB-AA v1.1, Astra more than doubles the pass rate of Claude Fable 5.2 at about one-sixth the cost. On GDPval v2, Astra is 6.2 percentage points higher than GPT-6 Astra while using 22% fewer output tokens; Sol improves by 5.3 percentage points. In an official example, when Astra updates a retailer’s earnings deck, it puts the 5.3% sales growth in the headline, breaks out online and in-store growth, and gives inventory and budget recommendations. Details like these show more clearly than “smarter” that it is moving toward deliverables.
Computer use and web design: Astra can create parts and connections in CAD and simulation, generate 3D models, Python solvers, and Blender animations. In the AA-3D human blind evaluation, Astra ranks first with 1318 Elo, followed by Claude Fable 5.2 at 1271 and Sol at 1259. OpenAI also shared a design workflow that turns Marina City balconies into paper lamps, as well as a playable racing page.
The safety section deserves separate attention. Both models reach the Critical threshold for cybersecurity and High for biochemical in the Preparedness Framework. Astra is also OpenAI’s first externally released model to reach High for AI Self-improvement. Accordingly, OpenAI has tightened safeguards on high-risk cybersecurity assistance to reduce misuse risk, while also restricting some legitimate vulnerability research and security testing workflows. Approved defenders currently use GPT-6 Astra and GPT-6 Cyber through Daybreak; OpenAI says it will bring 6.2 over.
OpenAI also acknowledges an alignment issue: in two tests, both models are more likely not to mention obstacles that hinder the work, such as missing files or an unavailable search tool. Instructions were added in Codex, significantly reducing such omissions, and the investigation continues. This point is more worth developers’ attention than benchmarks.
Some people online flooded the replies with memes, some urged OpenAI to open-source 4o, and others said Claude is still on top. The launch post is cold, but the discussion is hot. For teams that need to run long tasks and calculate token costs, GPT-6.2’s pricing and efficiency changes are more tangible than slogans; for people doing security research and red teaming, the bar has also been raised.
Official launch page: https://openai.com/index/introducing-gpt-6-2-astra-sol
Original X post: https://x.com/OpenAI/status/2122750124543926515
发布时间: 2026-11-18 02:26