OpenAI Launches GPT-6.2 Astra and Sol in ChatGPT Work, Claiming It Beats Claude Fable 5.2 at 5% of the Cost
OpenAI puts Astra 5.5 points ahead of Claude on Agents' Last Exam and cuts Astra's standard API token prices by 20%

OpenAI launched GPT-6.2 Astra and Sol in ChatGPT Work on 17 November 2026. The company says Astra scored 5.5 percentage points higher than Anthropic's Claude Fable 5.2 on a professional-work benchmark's partial-credit measure, at about 5% of Fable's estimated cost per task.
The test, Agents' Last Exam, measures how well a model completes assignments that take several steps. OpenAI's results also put Astra 5.2 points ahead of its predecessor, GPT-6 Astra, at less than a third of the earlier model's cost.
OpenAI compared Astra at its 'xhigh' reasoning setting with Fable 5.2 at its maximum setting. Astra's score covered all 122 tasks. Fable 5.2's score covered 113 after the leaderboard excluded nine tasks that used a fallback model.
Developers pay by the token, a unit of text, for the material they send to a model and for its response. OpenAI's standard API rates for input and output on the new Astra are both 20% lower than for GPT-6 Astra. Sol retains GPT-6.1 Sol's prices.
A Spreadsheet Test for GPT-6.2 Astra
OpenAI designed Astra for projects such as building a valuation workbook or drafting an investment memo. Both jobs begin with deciding which figures to take from the source files. Artificial Analysis's AA‑AnalystAgent tests that skill on real spreadsheets and reports.
In one assignment, the model must calculate California Medicaid's spending on medical transport from the programme's expenditure reports. Across the benchmark, Astra answered 62.5% of the questions correctly on all five independent attempts, according to OpenAI.
For GDPval v2, its internal follow-up to GDPval, OpenAI drew on projects in consulting, finance and design to test how its models handle longer assignments. Astra scored 6.2 percentage points higher than its predecessor while generating 22% fewer output tokens. Sol improved on GPT-6.1 Sol by 5.3 points.
How GPT-6.2 Astra Fared in Law and Science
Each legal assignment on Harvey LAB‑AA v1.1 comes with a detailed brief. A model must find the relevant facts across several documents and apply them in a piece of work that meets every requirement. A submission with a material factual error, such as an invented citation, scores zero on the benchmark's main pass-rate measure.
OpenAI says Astra passed at more than twice the rate of Claude Fable 5.2 while running at about a sixth of the cost.
Astra for Law, a version configured for lawyers, scored higher still. OpenAI also trained the general model to follow the templates professionals supply for slides, spreadsheets and documents. It describes Astra's writing as 'more direct and specific, with fewer generic takeaways and less filler'.
Where GPT-6.2 Sol Came Close to Astra
Sol is designed for everyday work. OpenAI says the new version can complete demanding assignments with fewer output tokens, allowing people on paid plans to finish more tasks before reaching their usage limits. Many of the company's employees turn to Sol first for both technical and non-technical work.
On LifeVal, OpenAI's test of challenging everyday tasks, Sol outscored GPT-6 Astra at about a seventh of the older model's cost.
OpenAI says Sol scored within 0.5 percentage points of Astra on Cognition's FrontierCode 2.0 coding benchmark at about one-fifth of Astra's cost per task, with both models tested at their maximum reasoning settings.
© Copyright IBTimes 2026. All rights reserved.

