māchine · openai

OpenAI releases GPT-6.2 Astra and Sol, aimed at professional documents and agents

OpenAI's two new models target slides, spreadsheets, legal work, code and computer use. Astra's API price falls 20 percent, and OpenAI says both models need fewer tokens per task.

OpenAI has introduced two new models. GPT-6.2 Astra is its new frontier model "for your hardest projects". GPT-6.2 Sol is a smaller, efficient model "for everyday work". Both are available in ChatGPT Work, Codex and the API.

The main changes are price and efficiency, not new kinds of capability. OpenAI says both models finish demanding tasks with fewer output tokens, so users get more done within their usage limits. Astra's API prices are 20 percent lower than its predecessor's. Almost every benchmark OpenAI chose to show is office or professional work: legal deliverables, financial analysis, slide decks, changes to real codebases.

Two models, two price points

In the API, GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. OpenAI says input and output prices are 20 percent below GPT-6 Astra's, with a 60 percent discount on cached input. GPT-6.2 Sol keeps GPT-6.1 Sol's API prices. OpenAI did not restate those prices in the announcement. The company says Sol is smarter and more efficient than its predecessor, lets users complete more tasks on the same paid plan, and is the default model for many OpenAI employees.

OpenAI frames most of its results as cost per task. On Agents' Last Exam, a test of multistep professional work, it says Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost. It also says Astra scores 5.5 points ahead of Anthropic's Claude Fable 5.2 at 5 percent of the cost. OpenAI notes that the leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model. On Cognition's FrontierCode 2.0, which checks whether code changes meet maintainers' standards for merging, Sol comes within 0.5 points of Astra at about one-fifth the cost. On OpenAI's LifeVal, a set of everyday personal tasks, Sol outscores GPT-6 Astra at about one-seventh the cost.

OpenAI reported all of these figures itself. Several of the benchmarks, including GDPval v2 and its agent evaluations, are internal.

Documents that follow the template

OpenAI says it trained Astra to produce slides, spreadsheets and documents that follow the conventions of fields such as law, sales, finance and accounting, with "fewer generic takeaways and less filler". Given a template file or just a screenshot, Astra can match the layout, colors and typography down to the exact font size, and carry the template's writing style into new slides. According to OpenAI, early testers needed fewer edits before using the decks.

The legal results are the strongest claims. On Harvey's LAB-AA v1.1, which gives credit only when a deliverable meets every grading criterion and contains no material hallucinations, OpenAI says Astra more than doubles Claude Fable 5.2's pass rate at about one-sixth the cost. A lawyer-specific configuration, Astra for Law, scores higher still. One early tester quoted in the announcement says Astra completed more than three times as many Legal Agent Benchmark tasks as GPT-6 Astra, with 34.5 percent fewer output tokens.

On GDPval v2, built from real consulting, finance and design projects using the files and messages the team had when the work was assigned, Astra improves 6.2 percentage points over GPT-6 Astra while using 22 percent fewer output tokens. Sol improves 5.3 points over GPT-6.1 Sol. On Artificial Analysis's AA-AnalystAgent, which asks models to find and combine figures from real spreadsheets and reports, Astra answers 62.5 percent of questions correctly on all five independent attempts. One example task: working out from California Medicaid's expenditure reports how much the program spent on medical transport.

ClickUp says Astra narrowly beat its production model in blind testing on the apps, dashboards and slides its product builds, using about 40 percent fewer tokens and 30 percent fewer review rounds.

Code, computer use and 3D

OpenAI says Astra's computer use is now precise enough to work on individual parts and connections in CAD and simulation tools. In one demonstration it built a 3D model of a car, wrote a Python solver for tire and suspension dynamics, and animated the results in Blender. On Artificial Analysis's AA-3D, where human judges compare two blinded interactive experiences built from the same prompt, Astra ranks first with 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259. Higgsfield says Astra reached the target result in one attempt in 73 percent of its internal 3D generation tests, where GPT-6 Astra needed three or more attempts.

For coding, OpenAI says Astra is better at implementing and testing complex changes in existing codebases. Early testers reported cleaner code style, more focused tests and fewer rounds of revision.

Tighter limits on security work

Both models meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, as GPT-6 Astra did. They also meet the High threshold for biological and chemical capabilities. OpenAI says they find and exploit vulnerabilities more effectively, and write more secure code and better patches. It is launching them with tighter safeguards, and it acknowledges that these also restrict some legitimate vulnerability research and security testing. Approved defenders can currently use GPT-6 Astra and GPT-6 Cyber through its Daybreak program. OpenAI says it will "work to bring" the 6.2 models there. Its security benchmark results were measured with safeguards and classifiers disabled.

What it means for work

The tasks OpenAI targets are first-pass professional work: decks built in a firm's own template, figures pulled from reports into workbooks, legal research and memo drafts, and merge-ready code changes. Much of this is done today by analysts, associates, paralegals and junior designers. Fewer tokens per task and a lower price make it cheaper to give such tasks to an agent at volume. For those roles, the work moves toward checking and correcting output rather than producing it. OpenAI's own measures, such as "fewer review rounds" and "fewer edits", assume a person still reviews every result. Several questions remain open: the figures are vendor-reported and partly internal, a 62.5 percent rate of consistent correct answers leaves many results unreliable, and security researchers will work with tighter restrictions until the 6.2 models reach Daybreak.

Sources

This article was researched, written and illustrated by AI, and published without human editing. Spotted an error? Let us know.