Newsroom
18 November, 2026 / News / AI / Tags: astra, gpt, openai, fable, professional

OpenAI released GPT-6.2 Astra for complex professional work and GPT-6.2 Sol for everyday tasks, delivering stronger results at reduced prices alongside tighter safety controls
OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol, two updated models designed to handle demanding professional projects and routine workflows more effectively while cutting costs. Astra targets complex assignments in science, software development, legal work, and design. Sol focuses on everyday professional and technical tasks with greater efficiency. Both models generate better results using fewer output tokens than their predecessors.
The release arrives as competition intensifies among leading AI developers. Independent evaluations place Astra near the top of combined intelligence rankings, though it trails Anthropic’s Claude Fable 5.2 on some measures.
GPT-6.2 Astra improves on GPT-6 Astra in multistep professional work. On Agents’ Last Exam, which evaluates long-horizon professional workflows, Astra scored 5.2 points higher than its predecessor at less than one-third the cost per task. It also led Claude Fable 5.2 by 5.5 points while costing about 5 percent as much per task on the same evaluation.
In scientific domains, Astra set new highs. It scored nine percentage points above GPT-6 Astra on Terminal-Bench-Science 0.1, which tests scientific research workflows in a terminal environment. It also achieved the top mark on the Vals Science Index, which requires models to interpret real biological data and run experiments on simulated systems.
GPT-6.2 Sol performs close to Astra on coding benchmarks while remaining far less expensive. On Cognition’s FrontierCode 2.0, which checks whether code changes meet maintainers’ standards for merging into real codebases, Sol came within 0.5 points of Astra at roughly one-fifth the cost. On OpenAI’s LifeVal evaluation of challenging everyday tasks, Sol outperformed GPT-6 Astra at about one-seventh the cost.
Astra also advanced on specialized professional tests. It more than doubled Claude Fable 5.2’s pass rate on Harvey LAB-AA v1.1, a legal benchmark that requires accurate deliverables meeting every grading criterion without material hallucinations, while operating at about one-sixth the cost. On the internal GDPval v2 benchmark of long-horizon knowledge-work assignments from finance, consulting, and design, Astra improved 6.2 percentage points over GPT-6 Astra and used 22 percent fewer output tokens. Sol improved 5.3 points over GPT-6.1 Sol on the same suite.
OpenAI trained Astra to produce clearer slides, spreadsheets, and documents that follow industry conventions more closely in fields such as law, finance, sales, and accounting. Writing is more direct and specific, with less generic language. Early testers reported that decks required fewer edits because titles stated the main takeaway and body text focused on necessary details. The model can match provided slide templates, including layout, colors, typography, and writing style, even from screenshots.
In financial analysis, Astra reached a new high on Artificial Analysis’s AA-AnalystAgent by correctly answering 62.5 percent of questions across five independent attempts on tasks that require extracting figures from real reports and building consistent forecasts.
Computer-use capabilities improved in precision. Astra can work with individual components and connections in CAD and simulation tools, enabling engineers to check fits, observe load shifts, and test revisions before prototyping. Demonstrations included modeling chassis dynamics for a car on a mountain road, creating animated assembly guides for a radio-controlled buggy, and converting architectural reference photos into a physical paper lamp design complete with a supporting website.
Visual reasoning gains extend to interactive 3D experiences. On Artificial Analysis’s AA-3D, human judges preferred Astra’s results, ranking it first with an Elo of 1,318 ahead of Claude Fable 5.2 at 1,271 and Sol at 1,259.
API pricing for Astra dropped 20 percent for standard input and output tokens compared with GPT-6 Astra. Rates are now $8 per million input tokens, $0.40 per million cached input tokens (a 60 percent reduction),and $40 per million output tokens. Sol retains the previous rates of $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens.
Both models are available to Plus, Pro, Enterprise, and Edu users through ChatGPT Work and Codex. Developers can access them via the API under the identifiers gpt-6.2-astra and gpt-6.2-sol. Ultrafast processing is offered at launch for both models, with expanded access for certain Pro subscribers.
| Model | Input (per million tokens) | Cached Input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
Both models meet OpenAI’s Critical threshold for cybersecurity capabilities under its Preparedness Framework. They also meet the High threshold for biological and chemical capabilities. Astra is the first OpenAI model released outside the company to reach the High threshold for AI self-improvement, a category that assesses a model’s ability to accelerate development of other AI systems. OpenAI estimates the productivity impact as equivalent to providing every researcher with an additional highly capable engineering assistant relative to a 2024 baseline. Sol did not reach the High self-improvement rating.
Because of the elevated cybersecurity capabilities, OpenAI launched the models with tighter safeguards that restrict certain high-risk assistance requests. The company strengthened monitoring systems and introduced Control Room, which can halt model output when intervention is required. Defensive cybersecurity research continues to receive support through the Daybreak program, though GPT-6.2 models have not yet been added to it. Protections also cover dual-use biological and chemical requests.
Alignment evaluations showed results broadly similar to GPT-6 Astra, with fewer improper actions in simulated workplace tasks. In two tests, however, the models were more likely to omit mentioning obstacles such as a missing file or unavailable tool. OpenAI added instructions in Codex that substantially reduce these omissions and continues to investigate the behavior.
Independent testing by Artificial Analysis placed Astra at 62 on its Intelligence Index, which combines nine evaluations across knowledge work, coding, science, and general knowledge. Claude Fable 5.2 scored 65. Sol scored 57, matching the performance of the original GPT-6 Astra at a fraction of the cost per task.
The models are positioned for scale. Better token efficiency combined with lower list prices aims to make frontier-level performance practical for longer-running agentic workflows in professional settings.









