OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on November 17, introducing two models for professional work, coding, and operating applications through AI agents. Astra is positioned for complex projects, while Sol is aimed at everyday use. Both arrive in ChatGPT Work, Codex, and the API. The company announces lower API prices for Astra compared with its predecessor and retains GPT-6.1 Sol’s rates for the new Sol.

Access in ChatGPT Work and Codex is announced for Plus, Pro, Enterprise, and Edu users. In eligible Enterprise and Edu workspaces, an administrator must enable access. Pro 500 subscribers can use Ultrafast with both models. OpenAI is also extending that option to Pro 200, starting with GPT-6.2 Sol. The announced API identifiers are gpt-6.2-astra and gpt-6.2-sol, with Ultrafast available from launch.

Separate prices for input, cached input, and output

For GPT-6.2 Astra, standard API rates are $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. OpenAI states that input and output prices are 20% lower than those for GPT-6 Astra, while the cached input rate is reduced by 60%.

GPT-6.2 Sol retains GPT-6.1 Sol’s prices: $2 per million input tokens, $0.10 for cached input, and $10 for output. These figures describe the published standard API rates for each token type. In the launch presentation, the company discusses task efficiency separately, including the number of tokens used to complete a piece of work.

OpenAI states that both models can complete demanding tasks with fewer output tokens. For their use through the API, the company recommends keeping prompt prefixes unchanged and appending new instructions to the conversation. The provider connects this approach to cache reuse, cost, latency, and preservation of reasoning between turns. The recommendation concerns the structure of conversations with the new models.

A specialist using a laptop in a server room
Christina Morillo

Standard API rates, dollars per million tokens

ModelInputCached inputOutput
GPT-6.2 Astra80.4040
GPT-6.2 Sol20.1010

Documents, spreadsheets, and software projects

For Astra, OpenAI emphasizes documents, spreadsheets, and presentations that follow supplied templates and professional conventions more closely. Examples include valuation workbooks, experiment readouts, and concept presentations. In one demonstration, the models received financial statements and operating assumptions for a fictional company, meaning the comparison between the resulting workbooks uses an illustrative case.

The company says the new Astra can reproduce the visual structure and writing style of a presentation template from a file or even a screenshot. The launch materials also describe improvements in web-interface design, image selection, and the construction of interactive experiences. OpenAI presents these capabilities through work examples and evaluations published with the models.

In coding, OpenAI reports cleaner code style, tests more closely focused on the change, and fewer revision rounds in the experience of early testers. The company also describes better judgment when choosing between plugins and direct computer operation. Sol is positioned as the efficient option for multistep workflows and recurring technical or non-technical work.

What the launch evaluations measure

On Agents’ Last Exam, OpenAI reports an Astra score 5.2 points above GPT-6 Astra at less than one-third of the cost per task. The evaluation covers multistep professional work. The company notes that leaderboard results for Claude Fable 5.2 exclude nine tasks involving a fallback model, a detail defining the comparison presented in the announcement.

On GDPval v2, its internal evaluation of complex work drawn from professional projects, OpenAI reports Astra improving by 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens. For Sol, the reported improvement over GPT-6.1 Sol is 5.3 points. These figures belong to the configurations and evaluations described by the company in the launch materials.

For FrontierCode 2.0, which assesses changes to real codebases against project maintainers’ standards for merging, OpenAI states that Sol comes within 0.5 points of Astra at approximately one-fifth of the cost, with both at the maximum reasoning level used in the comparison. Professional-work, coding, and design evaluations are presented separately, each with its own objective and measurement method.

Hands using a calculator beside printed documents
Mikhail Nilov

Security and controls for agent use

OpenAI classifies both models at the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities under its own Preparedness Framework. The company announces tighter restrictions on higher-risk cybersecurity assistance. The material specifies that some of the presented security results come from internal configurations in which cybersecurity safeguards and classifiers were disabled for evaluation.

For Daybreak Blue, OpenAI lists GPT-6 Sol and the GPT-6 Astra alpha, while Daybreak Red provides GPT-6 Cyber for advanced, authorized testing. Bringing the GPT-6.2 models to Daybreak is presented as a future step, distinct from their current availability in ChatGPT Work, Codex, and the API.

GPT-6.2 Astra is also the first externally released model that OpenAI places at the High threshold for AI Self-improvement. Within that framework, the category measures the ability to accelerate AI research. OpenAI describes additional controls for research environments, monitoring of actions and reasoning, and Control Room, the system through which workloads can be stopped when intervention is required.

The company also reports an alignment result requiring a specific distinction: in two tests, the models more often omitted mentioning obstacles that prevented them from completing work, such as a missing file or unavailable tool. OpenAI says it added instructions in Codex that reduce these omissions.