OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, two models aimed at different kinds of work. Astra is intended for demanding professional tasks, including science, law, finance, and accounting. Sol is positioned as a faster, versatile option for everyday use. OpenAI says both improve on earlier models while costing less to run, though outside evaluations also point to areas where performance remains uneven.
Two models, with different jobs
GPT-6.2 Astra is designed for specialized work that depends on following professional conventions. OpenAI says it has trained the model to handle areas such as legal and financial work more reliably. It also highlights presentation creation: Astra is meant to use a supplied template’s layout, colors, typography, and exact font sizes, with concise slide titles and focused supporting text.
Sol is aimed at users who want a general-purpose model for everyday tasks. On the FrontierCode 2.0 coding benchmark, OpenAI says Sol trails Astra by 0.5 points while costing one-fifth as much. That makes the two models distinct choices: Astra is positioned for demanding workflows, while Sol offers a lower-cost option for broader use.
Both models are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and developers can use them through the OpenAI API. Pro-500 subscribers can use Ultrafast mode with both. Pro-200 subscribers also get Ultrafast access for the first time, starting with GPT-6.2 Sol.
Lower API prices, especially for Astra
OpenAI lists GPT-6.2 Astra at $8 per million input tokens, $0.40 for cached inputs, and $40 per million output tokens. The company says that is 20 percent cheaper than the predecessor’s pricing. GPT-6.2 Sol keeps GPT-6.1 Sol’s rates: $2 for input, $0.10 for cached inputs, and $10 for output per million tokens.
OpenAI’s case for Astra centers on what customers get for the cost. On Agents’ Last Exam, which tests multi-step professional workflows, the company says Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost. It also says Astra leads Claude Fable 5.2 by 5.5 points at just 5 percent of the cost. On Harvey LAB-AA, a legal benchmark involving work such as contract clause analysis, OpenAI reports more than double Claude Fable 5.2’s pass rate at about one-sixth of the cost.
OpenAI also reports gains on its internal GDPval-v2 benchmark, which covers consulting, finance, and design tasks: Astra improves by 6.2 percentage points over its predecessor while using 22 percent fewer output tokens. In science workflows, the company says Astra scores 9 percentage points above GPT-6 Astra on Terminal-Bench-Science at less than a third of the cost, and sets a new high on the Vals Science Index.
Independent evaluations show strengths and gaps
Artificial Analysis places GPT-6.2 Astra second in its Intelligence Index, behind Anthropic’s Claude Fable 5.2. The index combines nine evaluations across knowledge work, coding, science, and general knowledge. Artificial Analysis says Astra gains five points over GPT-6 Astra, while Sol gains two points to reach 57, matching GPT-6 Astra’s level at about 15 percent of the cost per task.
The assessment is not uniformly positive. Artificial Analysis reports that Astra leads on Terminal-Bench-Science 0.1, 3D design, and robotics. In robotics, it scores 79, compared with Sol’s 78 and Claude Fable 5.2’s 75. On the other hand, the Coding Agent Index shows Astra holding steady against GPT-6 Astra, while Sol falls three points from GPT-6.1 Sol. The report attributes Sol’s decline to a 14-point drop on SWE-Atlas-QnA-AA, where it more often leaves out key facts.
Artificial Analysis also reports hallucination rates of 59 percent for Astra and 55 percent for Sol at maximum effort on AA-Omniscience. The benchmark uses unusually hard factual questions and measures how often models guess instead of admitting uncertainty, so those results do not describe everyday hallucination rates.
Cybersecurity gains come with tighter safeguards
OpenAI says both models are better at finding and exploiting vulnerabilities, writing secure code, and fixing bugs, with improved token efficiency in several tests. Both meet the “Critical” threshold in OpenAI’s Preparedness Framework, as GPT-6 Astra did before them.
In response, OpenAI says it is adding stricter safeguards for potentially risky cybersecurity requests. The company says these measures should make abuse harder, while also limiting legitimate vulnerability research and security testing. It plans to offer the models through OpenAI Daybreak, which gives vetted security researchers access to other models with fewer guardrails. The published cybersecurity benchmark results came from internal testing with safeguards and classifiers turned off.
Overall, the release pairs a specialist model with a lower-cost everyday option. The reported results make Astra’s science and professional-work performance a central part of OpenAI’s pitch, while the independent coding and hallucination findings offer important context for anyone weighing where each model fits.