OpenAI launches GPT-6.2 Astra and Sol, cuts API prices 20%
OpenAI ships GPT-6.2 Astra and Sol for professional and office work, cutting Astra API prices 20% and claiming benchmark leads over Claude Fable 5.2 at a fraction of the cost.

Why it matters
- GPT-6.2 Astra API pricing: $8 input, $0.40 cached input, $40 output per million tokens — 20% lower than GPT-6 Astra.
- On Agents' Last Exam, OpenAI says GPT-6.2 Astra scored 5.5 points ahead of Claude Fable 5.2 at 5% of the cost.
- On Harvey LAB-AA v1.1, OpenAI claims GPT-6.2 Astra more than doubled Claude Fable 5.2's pass rate at about one-sixth of the cost.
- GPT-6.2 Sol came within 0.5 points of Astra on Cognition's FrontierCode 2.0 at about one-fifth of the cost.
- GPT-6.2 Astra answered 62.5% of questions correctly on California Medicaid expenditure reports in the AA-AnalystAgent benchmark.
OpenAI has launched two new models, GPT-6.2 Astra and GPT-6.2 Sol, and cut Astra's API input and output prices by 20% compared with GPT-6 Astra. The company priced the model at USD $8 for input, USD $0.40 for cached input and USD $40 for output per million tokens. GPT-6.2 Sol keeps the same API token prices as GPT-6.1 Sol.
The split positioning is straightforward. GPT-6.2 Astra targets complex professional work such as financial modelling, legal analysis and scientific research. GPT-6.2 Sol handles multi-step everyday tasks including writing, coding and project planning. Both models roll out in ChatGPT Work and Codex and will also be available through the API. The launch matters because OpenAI is competing on cost per task, not only raw benchmark scores — a shift that directly affects what enterprise customers pay at scale.
What do the benchmarks show?
Every number below comes from OpenAI itself, and the company has a clear incentive to frame comparisons favorably. On Agents' Last Exam, a test of multi-step professional workflows, OpenAI said GPT-6.2 Astra scored 5.2 points higher than GPT-6 Astra at less than one-third of the cost. Against Anthropic's Claude Fable 5.2 on the same test, OpenAI claimed Astra scored 5.5 points ahead at 5% of the cost.
One detail deserves scrutiny: OpenAI added that published leaderboard results for Fable 5.2 excluded nine tasks that used a fallback model. That caveat cuts both ways. It may explain a gap in the rival's results, or it may signal that leaderboard hygiene across the industry is uneven. Neither company's numbers here are independently verified. The takeaway in plain words: OpenAI claims a large lead, but the comparison rests on contested test conditions.
For scientific work, OpenAI said Astra reached new highs on Terminal-Bench-Science 0.1, which tests whether a model can carry out scientific work in a terminal environment, and on the Vals Science Index. On Terminal-Bench-Science 0.1, Astra scored nine percentage points higher than GPT-6 Astra at less than one-third of the cost, according to the company.
Legal and business workflows got their own comparisons. On Harvey LAB-AA v1.1 — a benchmark using legal assignments to assess whether a model can identify relevant facts and produce an accurate deliverable — OpenAI said GPT-6.2 Astra more than doubled Claude Fable 5.2's pass rate at about one-sixth of the cost. On GDPval v2, an internal OpenAI benchmark built around long-horizon assignments from consulting, finance, design and other sectors, Astra improved by 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens. GPT-6.2 Sol improved by 5.3 points over GPT-6.1 Sol on the same benchmark, the company said. Internal benchmarks invite the obvious caveat: the test author controls the ruler.
Why is cost the centerpiece?
OpenAI said both models are designed to produce stronger results while using fewer output tokens, which could lower the cost of each task for customers working within usage limits. Note the conditional — "could," not "will." Actual savings depend on whether real-world tasks compress as neatly as benchmark tasks do.
The pricing details are concrete:
- GPT-6.2 Astra input: USD $8 per million tokens
- Cached input: USD $0.40 per million tokens (a 60% discount on cached input)
- Output: USD $40 per million tokens
- Input and output prices are 20% lower than GPT-6 Astra's
- GPT-6.2 Sol pricing is unchanged from GPT-6.1 Sol
OpenAI said several benchmark results showed the new models close to or ahead of other leading systems at a lower cost per task. That framing reflects the broader competitive reality: model developers are now fighting over the balance between quality and operating expense, because enterprise buyers measure both. In plain words, the pitch is no longer "our model is smarter" but "our model is smarter per dollar."
What is GPT-6.2 Sol for?
OpenAI described GPT-6.2 Sol as its standard model for routine work. It targets tasks that take several steps: converting meeting notes into project plans, tracing software bugs and handling everyday personal organization. In Cognition's FrontierCode 2.0 benchmark, which tests whether changes to real codebases meet maintainers' standards for merging, GPT-6.2 Sol came within 0.5 points of GPT-6.2 Astra at about one-fifth of the cost when both ran at maximum settings, OpenAI said.
OpenAI also cited LifeVal, an internal evaluation measuring performance on difficult everyday tasks tailored to individual needs. On it, GPT-6.2 Sol outscored GPT-6 Astra at about one-seventh of the cost, according to the company. The examples OpenAI gave were specific: shortlisting kindergartens after a family moves, finding a home that fits a family's budget and commute, and planning meals for a parent adapting to a new diagnosis. These are usefully concrete — and they are OpenAI's own chosen illustrations, not independent user reports. The takeaway: Sol is positioned as the volume model for daily work, with Astra reserved for harder projects.
How does the output quality claim hold up?
OpenAI said it trained GPT-6.2 Astra to produce slides, spreadsheets and documents that better match professional conventions in law, sales, finance and accounting. The company also said the model's writing is more direct and specific, with less filler and fewer generic conclusions. These are qualitative claims with no published metric attached; readers should treat them as vendor description until third-party testing appears.
The one externally constructed test OpenAI cited is Artificial Analysis's AA-AnalystAgent benchmark, which tests work on real spreadsheets and reports. In one task using California Medicaid expenditure reports, GPT-6.2 Astra answered 62.5% of questions correctly across five independent attempts, according to the company. A 62.5% success rate on real government financial data is a solid result — and also a reminder that "professional-grade" still means the model gets roughly one question in three wrong. In plain words: the model is improving on real documents, but the error rate on this test remains far from negligible.
What is the stakes for the market?
The launch underscores a strategic shift across the industry: developers increasingly target practical office and professional workflows rather than only general-purpose chatbot use. OpenAI explicitly divided its lineup along those lines, calling GPT-6.2 Astra its model for harder projects and GPT-6.2 Sol the model for a broader set of daily analytical, writing and coding tasks.
The competitive claims against Claude Fable 5.2, paired with aggressive price cuts, signal that OpenAI wants to win enterprise workloads on unit economics. The unresolved question is durability: if token efficiency and 20% price cuts become table stakes, margins across the industry compress further. For now, buyers evaluating the switch should ask for one thing the launch materials do not provide — independent reproduction of the benchmark numbers.
Source: itbrief.com.au
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
722 articles