OpenAI just dropped GPT-6.2 Astra and Sol, and the headline isn't just "better"—it's "better and cheaper by an order of magnitude." This is the kind of release that changes deployment math for anyone running agents or professional workflows at scale.
GPT-6.2 Astra is the new frontier model, while GPT-6.2 Sol slots in as the everyday workhorse. Both deliver what OpenAI calls "better results on demanding tasks with fewer output tokens," which translates directly to doing more within your usage caps and burning less budget per task.
The cost story is wild
Let's start with the number that made me sit up: on Agents' Last Exam—a benchmark testing multistep professional workflows—GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost. It also beats Claude Fable 5.2 by 5.5 points at just 5% of the cost.
In the API, Astra's input and output token prices dropped 20% compared to GPT-6 Astra, with cached input at $0.40 per million tokens (down from the implied $2 baseline, given the 60% discount). That's $8 input, $0.40 cached input, $40 output per million tokens. Sol keeps GPT-6.1 Sol's pricing unchanged but delivers measurably better performance.
This isn't incremental. The efficiency gains let you run workflows that were cost-prohibitive last month.
Professional outputs that actually look professional
OpenAI trained GPT-6.2 Astra specifically to generate slides, spreadsheets, and documents that match industry conventions in law, sales, finance, and accounting. The examples in the announcement show Astra building Excel valuation workbooks with proper reconciliation across three financial statements, sensitivity tables positioned next to revenue bridges, and layouts that make the analytical narrative immediately scannable.
On Harvey LAB-AA v1.1—a legal benchmark requiring models to extract facts from source documents, apply them to detailed briefs, and produce deliverables meeting every requirement—Astra more than doubles Claude Fable 5.2's pass rate at about one-sixth the cost. There's also a specialized Astra for Law configuration that scores even higher.
The deck design improvements are surprisingly specific. Astra can now match slide templates down to exact font sizes, carry writing style into new slides, and produce what early testers describe as "easier to use directly with fewer edits." The comparison slides show GPT-6 Astra producing generic mockups while 6.2 Astra delivers a cohesive illustrated field guide with team-specific jokes, consistent layouts, and production-ready vector artwork with proposed dimensions.
This kind of template fidelity matters enormously in enterprises where brand compliance and professional polish aren't nice-to-haves.
Sol closes the gap at a fraction of the cost
GPT-6.2 Sol is what many OpenAI employees reportedly use as their default for both technical and non-technical work. On Cognition's FrontierCode 2.0—which tests whether code changes to real codebases meet maintainer standards for merging—Sol comes within 0.5 points of Astra at about one-fifth the cost.
On LifeVal, OpenAI's recently released benchmark for challenging everyday tasks (shortlisting kindergartens after a move, finding homes fitting budget and commute constraints, meal planning around new diagnoses), GPT-6.2 Sol outscores GPT-6 Astra at roughly one-seventh the cost.
That's the kind of price-performance ratio that makes you rethink your model routing strategy.
Computer use, CAD, and 3D workflows
GPT-6.2 Astra brings meaningfully improved computer use, working down to individual parts and connections in CAD and simulation tools. The examples include:
- Visualizing chassis dynamics of a car on a mountain road, creating the 3D model and writing a Python solver to simulate tire-suspension interactions, then animating results in Blender
- Modeling a 1:8-scale nitro RC buggy kit part by part from public photos and build manuals, producing animated assembly guides and illustrated parts catalogs
- Turning Chicago's Marina City balconies into a CAD-designed paper lamp, with a generated site for creating your own
On Artificial Analysis's AA-3D benchmark—where human judges compare blinded 3D experiences built from identical prompts—GPT-6.2 Astra ranks first at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259.
The computer-use precision extends to web design, with Astra choosing layouts and imagery that give sites distinctive looks and building interactive elements (vehicle configurators, explorable 3D environments, playable pinball tables) around actual site goals rather than generic templates.
Code quality and workflow intelligence
Early testers report GPT-6.2 Astra produces cleaner code style with more focused tests and fewer revision rounds. The model has what OpenAI describes as "better judgment about how to execute a task," including matching tools and effort to the workflow and knowing when to use plugins versus computer use.
This operational intelligence—deciding how to solve something, not just solving it—is the kind of capability that makes agentic workflows less brittle in production.
Science benchmarks and the cybersecurity threshold
OpenAI claims GPT-6.2 Astra is "the best model in the world for scientists," setting new highs on Terminal-Bench-Science 0.1 (scientific research workflows in a terminal) with a 9-percentage-point improvement over GPT-6 Astra at under one-third the cost. It also leads on the Vals Science Index, which tests biological data interpretation and experimental inference.
Both 6.2 Astra and Sol cross the Critical threshold for cybersecurity under OpenAI's Preparedness Framework. They find and exploit vulnerabilities more effectively, write more secure code, and produce better patches with substantial token-efficiency gains. OpenAI is launching them with tighter safeguards for higher-risk cybersecurity assistance, which restricts some legitimate vulnerability research and security testing workflows.
Defenders can access less-restricted versions through OpenAI Daybreak.
What this means for deployment
The combination of better performance and dramatically lower per-task costs reshapes the economics of LLM deployment. Workflows that were marginal at GPT-6 pricing become viable. Agentic systems that burned through token budgets too quickly can now run longer.
The professional output quality—actual template matching, industry-convention awareness, cleaner deliverables—addresses one of the persistent friction points in enterprise adoption: the "AI slop" problem where outputs require so much human editing that the productivity gains evaporate.
If the benchmark numbers hold up in production, this is the kind of release that makes you revisit every "too expensive to automate" decision from the last six months.