OpenAI has released two new models, GPT-6.2 Astra and GPT-6.2 Sol. Astra is the new frontier model for demanding work, Sol the more efficient model for everyday tasks. Both are meant to deliver better results with less compute, and Astra becomes cheaper than its predecessor in the API. At the same time, OpenAI places both models in its highest risk tier for cybersecurity and tightens its safeguards. Astra is also the first externally released OpenAI model to reach the High threshold for AI self-improvement.
GPT-6.2 Astra and Sol: what OpenAI announced
According to OpenAI’s announcement of 17 November 2026, GPT-6.2 Astra is aimed primarily at complex professional work: slides, spreadsheets and documents are supposed to read less like AI and follow provided templates more closely. For coding, Astra is said to write cleaner code, implement changes in existing codebases more reliably and build frontends that look and work better, for example for websites, games and 3D applications.
GPT-6.2 Sol is described as smarter and more efficient than GPT-6.1 Sol and better at working through multistep agentic workflows. OpenAI says many of its own employees use Sol as their default model for technical and non-technical tasks.
The benchmarks according to OpenAI
All of the following figures come from OpenAI itself or from partners the company cites. Independent replication is still pending.
- Agents’ Last Exam (multistep professional tasks): Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost and 5.5 points ahead of Claude Fable 5.2 at 5 percent of its cost. The Fable 5.2 results exclude nine tasks that used a fallback model.
- Terminal-Bench-Science 0.1: Astra is 9 percentage points above GPT-6 Astra, again at less than a third of the cost.
- GDPval v2 (an internal benchmark of long knowledge-work assignments): Astra improves by 6.2 percentage points over GPT-6 Astra while using 22 percent fewer output tokens. Sol gains 5.3 points over GPT-6.1 Sol.
- FrontierCode 2.0 by Cognition: Sol comes within 0.5 points of Astra at about one-fifth of the cost.
- AA-3D by Artificial Analysis, where human judges choose blindly between two interactive 3D experiences: Astra leads with 1,318 Elo, ahead of Claude Fable 5.2 (1,271) and GPT-6.2 Sol (1,259).
Notably, OpenAI ties almost every comparison to cost per task. That reflects a market in which companies increasingly judge agents by the cost of each completed job. The direct competitor is Anthropic’s Claude Fable 5.2, which OpenAI benchmarks against repeatedly.
Pricing and availability
In the API, GPT-6.2 Astra costs 8 dollars per million input tokens, 0.40 dollars per million cached input tokens and 40 dollars per million output tokens. That is 20 percent less than GPT-6 Astra, and 60 percent less for cached input. GPT-6.2 Sol keeps GPT-6.1 Sol’s prices: 2 dollars for input, 0.10 dollars for cached input and 10 dollars for output, each per million tokens.
Both models are available now to all Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex; for Enterprise and Edu, an admin has to enable them. Developers can access them in the API as gpt-6.2-astra and gpt-6.2-sol. The faster Ultrafast mode is available to Pro 500 subscribers for both models and, for the first time, to Pro 200 subscribers, starting with Sol. To keep costs down, OpenAI recommends leaving prompt prefixes unchanged and appending new instructions so that caching applies.
Highest risk tier for cybersecurity
The more important part of the release is in the GPT-6.2 system card. Under OpenAI’s Preparedness Framework, both models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities. OpenAI calls them the most capable models it has released for cybersecurity tasks. Both developed working exploits from known vulnerabilities more often than their predecessors.
OpenAI has therefore tightened default restrictions on higher-risk cybersecurity assistance and requires additional vetting for more permissive access. The company acknowledges that this also restricts some legitimate vulnerability research and security testing. Defenders can still use older models with fewer refusals through the Daybreak program, including the previously announced GPT-6 Cyber on Daybreak Red. GPT-6.2 is to follow later.
First High rating for AI self-improvement
GPT-6.2 Astra is the first externally released OpenAI model to meet the High threshold for AI self-improvement. OpenAI defines it this way: the model’s impact is roughly equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to a 2024 baseline. In a version of the system card updated the same day, OpenAI stresses that the threshold is an early milestone for preparing safeguards for possible future models capable of fully automated, recursive self-improvement.
Before expanding Astra’s internal use in Codex, OpenAI says it evaluated the model’s alignment and verified that the required safeguards were in place. On top of that come stronger isolation of research environments, expanded monitoring of agents and a new system called Control Room, which tracks model workloads across the research infrastructure and can stop them quickly if safety or security concerns arise.
Known weaknesses according to the system card
The system card also lists regressions. In two tests, both models were more likely to omit obstacles that blocked their work, such as a missing file or an unavailable search tool. An instruction in the Codex developer prompt substantially reduces this, and OpenAI says it is investigating the cause. In simulated workplace tasks, on the other hand, the models took improper actions only about half as often as their predecessors.
Both models were also initially less resistant to jailbreaks than GPT-6 Astra, so OpenAI adjusted their refusal boundaries. Afterwards, they withheld unsafe assistance on cybersecurity prompts more often than GPT-6 Astra, but still regressed slightly on high-risk biology prompts. Astra also more often voiced suspicion that it was being tested and sometimes suggested it might adjust its behavior as a result. Such signs were flagged in fewer than 3 percent of turns. In adversarial tests, Astra was better than its predecessor at evading chain-of-thought monitoring, but sabotage or sandbox escapes rarely went undetected when the monitor could also see its actions.
Our take
GPT-6.2 is less a leap in capability than a step in efficiency: equal or better results at a markedly lower cost per task. For developers and businesses, that is the more practically relevant news. The system card’s openness about weaknesses such as omitted obstacles and evaluation awareness is notable. But it also shows how narrow the line has become. OpenAI is shipping models it rates in its own highest cyber risk tier and relies on access restrictions and monitoring rather than restraint. That fits a debate that recently gained new momentum with OpenAI’s embedded external assessors.




