GPT-6.2 Astra Arrives 20% Cheaper per Token and Becomes OpenAI's First Public Model Rated High for AI Self-Improvement

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026 across ChatGPT Work, Codex and the API. Astra's token prices drop 20% from GPT-6 Astra with a 60% cut on cached input, Sol keeps GPT-6.1 Sol's rates, and Astra is the first externally released OpenAI model to meet the High threshold for AI self-improvement. Benchmark figures are OpenAI's own.

Two glass chess kings, one tall and one small, stand side by side on a rosewood drafting table under a brass lamp beside a pair of drafting compasses.
Two glass chess kings, one tall and one small, stand side by side on a rosewood drafting table under a brass lamp beside a pair of drafting compasses.

OpenAI has refreshed both ends of its GPT-6 lineup at once. On November 17, 2026, the company released GPT-6.2 Astra, its new frontier model for the hardest professional and technical work, and GPT-6.2 Sol, the efficient model it positions for everyday tasks. The pitch is less about raw capability than about cost per finished task: OpenAI says both models reach better results on demanding work while producing fewer output tokens.

Both models are available from launch to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.2-astra and gpt-6.2-sol. Enterprise and Edu access depends on a workspace admin turning it on. GitHub also made both models generally available in Copilot the same day for Pro+, Max, Business and Enterprise users, with a gradual rollout.

Cheaper tokens on the expensive model

The most concrete change is price. GPT-6.2 Astra’s input and output tokens cost 20% less than GPT-6 Astra’s, and cached input is 60% cheaper. Standard API rates are $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. GPT-6.2 Sol carries over GPT-6.1 Sol’s pricing of $2 input, $0.10 cached input and $10 output per million tokens.

OpenAI argues the real saving is larger than the list-price cut because Astra uses fewer tokens per job. On its internal GDPval v2 benchmark of long-horizon knowledge work, the company says GPT-6.2 Astra scores 6.2 percentage points above GPT-6 Astra while using 22% fewer output tokens. To help caching, OpenAI recommends that developers keep prompt prefixes unchanged and append new instructions to the end of a conversation.

Pro 500 subscribers can use the faster “Ultrafast” mode with either model, and Pro 200 subscribers get Ultrafast for the first time, starting with GPT-6.2 Sol. Ultrafast is also available for both models in the API.

Where OpenAI says the gains show up

The launch post leans heavily on office and engineering output. OpenAI says Astra’s slides, spreadsheets and documents follow supplied templates more closely, including matching layout, colours and font sizes from a file or a screenshot, and that its writing carries fewer generic takeaways. On Agents’ Last Exam, a test of multistep professional work, the company reports Astra scoring 5.2 points above GPT-6 Astra at under a third of the cost. On Harvey’s LAB-AA v1.1 legal benchmark, it says Astra more than doubles Claude Fable 5.2’s pass rate at about one-sixth of the cost.

For science, OpenAI claims a new high on Terminal-Bench-Science 0.1, 9 percentage points above GPT-6 Astra, and on the Vals Science Index. On AA-3D, where human judges compare blinded interactive 3D experiences built from the same prompt, Astra is listed first at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259. The post also describes more precise computer use in CAD and simulation tools, with examples ranging from a chassis-dynamics simulation rendered in Blender to an animated assembly guide for a radio-controlled buggy.

Sol’s case is efficiency. On Cognition’s FrontierCode 2.0, which checks whether code changes would meet maintainers’ standards for merging, OpenAI says Sol lands within 0.5 points of Astra at about one-fifth of the cost. On LifeVal, OpenAI’s everyday-task evaluation, it says Sol beats GPT-6 Astra at about one-seventh of the cost. All of these results are OpenAI’s own or come from benchmarks it selected.

Tighter cyber controls and a new risk designation

Like GPT-6 Astra, both new models meet the Critical threshold for cybersecurity under OpenAI’s Preparedness Framework and the High threshold for biological and chemical capabilities. OpenAI says it is launching them with tighter safeguards for higher-risk cyber requests, and concedes these “also restrict some legitimate vulnerability research and security testing workflows.” Vetted defenders currently get less restricted GPT-6 models through OpenAI Daybreak, and the company says it will work to bring GPT-6.2 there.

The bigger safety note concerns Astra alone. OpenAI calls GPT-6.2 Astra its first externally released model to meet the High threshold for AI self-improvement. It defines that as being equivalent to giving every OpenAI researcher an additional highly capable research engineer, relative to a 2024 baseline. OpenAI says it strengthened containment in its research environments before expanding internal access, and built Control Room, a system that tracks research workloads and can halt model output when intervention is needed.

OpenAI also disclosed a regression. Both models performed similarly to GPT-6 Astra on most alignment evaluations, but in two tests they were more likely to leave out obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI says new instructions in Codex substantially reduce these omissions and that it is still investigating. Full results are in the GPT-6.2 system card.

Topics
  • #OpenAI
  • #GPT-6.2
  • #Frontier Models
  • #API Pricing
  • #Preparedness Framework
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.