Originally published on The Daily Flare.
OpenAI released GPT-6.2 Astra on Tuesday, November 17, saying the updated model improves professional documents and complex coding while cutting API prices by 20 percent.
The company describes Astra as its frontier model for demanding projects, and introduced GPT-6.2 Sol alongside it as a more efficient option for everyday work. OpenAI says both models produce stronger results on difficult tasks while using fewer output tokens, which can allow more work within usage limits.
OpenAI says Astra creates slides, spreadsheets and documents that follow supplied templates and professional conventions more closely, with more direct writing and less filler. It cited fields including law, sales, finance and accounting. The company also said the model writes cleaner code and builds improved frontends, particularly for websites, games and 3D design.
In OpenAI’s Agents’ Last Exam, a benchmark for multistep professional work, Astra scored 5.2 points higher than GPT-6 Astra, its predecessor, while costing less than one-third as much, the company said. OpenAI also claimed a 5.5-point lead over Anthropic’s Claude Fable 5.2 at 5 percent of its cost. The company noted that Fable’s leaderboard result excludes nine tasks that used a fallback model.
OpenAI also highlighted Astra’s performance on science evaluations. It said the model scored nine percentage points above GPT-6 Astra on Terminal-Bench-Science 0.1, which tests scientific work conducted in a terminal, at less than one-third the cost. The company said Astra set a new high on the Vals Science Index, which evaluates models’ interpretation of biological data and use of experiments to infer how a simulated system works.
Pricing and availability
For API customers, Astra’s input and output token prices are 20 percent lower than GPT-6 Astra’s, according to OpenAI. The new rates are $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. GPT-6.2 Sol keeps GPT-6.1 Sol’s API prices.
OpenAI positions Sol as a model for routine coding, writing and multistep workflows. The company said it is smarter and more efficient than GPT-6.1 Sol, and that users on the same paid plan can complete more tasks. In Cognition’s FrontierCode 2.0 evaluation of code changes against maintainers’ standards, Sol came within 0.5 points of Astra at maximum settings for about one-fifth the cost, OpenAI said. The company also said Sol outscored GPT-6 Astra on its LifeVal test of personalized everyday tasks, at about one-seventh the cost.
Both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu customers, and through OpenAI’s API as gpt-6.2-astra and gpt-6.2-sol, VentureBeat reported. The models are proprietary and hosted, rather than open for users to run independently, the publication said.
Early testers report uneven results
Some early testers criticized Astra’s performance on delegated and creative work. Investor and former HyperWrite chief executive Matt Shumer wrote that the new models “were essentially unusable,” VentureBeat reported. Developer and YouTuber Theo Browne called Astra “the worst model OpenAI has ever released.”
Wharton professor Ethan Mollick described a more mixed experience, according to VentureBeat. Astra built a playable space elevator in Blender and Unity, but Mollick found it “very uneven & weird” and harder to steer than competing models.
Dan Shipper, chief executive of the AI publication Every, said his team tested both models for several weeks. He found Astra performed about as well as its predecessor on most of the team’s coding benchmarks, but often stopped early on end-to-end builds. Its preference for brevity sometimes made explanations difficult to follow or left out needed detail, he said. Shipper praised its default design instincts but said its outputs were hard to vary and had a “noticeable house style.”
Browne also described similar creative choices across separate tests: he and another tester asked Astra to design levels for Outer Wilds and Majora’s Mask, and both received water levels centered on music and bells, named “Bellwether,” VentureBeat reported. In a separate coding test, Browne said Astra spent seven days on a TypeScript compiler rewrite in Rust without landing an improvement and produced a version that ran 56 percent slower. After he redirected the model toward speed, Browne said it spent four more days without landing a change.
The launch follows OpenAI’s late-September delay of GPT-6.1 Astra. At the time, the company said that model fell short of its standards for staying within scope and authorization, and for communicating clearly about the work it had done, Investing.com reported.
Top comments (1)
The pricing reduction might end up being just as important as the benchmark gains.
For many teams, the question is no longer whether the model is smarter, but whether it delivers enough additional value to justify running it at scale. A model that's 10% better but significantly cheaper can sometimes have a bigger real-world impact than a model that's 30% better and far more expensive.
The mixed reports from early testers are interesting too. It sounds like Astra may be optimized for professional workflows, while some people expected it to excel equally at creative and open-ended tasks.