- OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, targeting professional and everyday tasks.
- The company said the models improve performance while using fewer tokens, and it cut Astra’s API prices for input and output tokens by 20%.
- Although the models posted strong benchmark results, several early testers criticized their writing, coding and design performance.
OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, with Astra aimed at complex professional tasks and Sol designed for everyday work. The release matters for users because OpenAI says both models deliver higher performance with fewer tokens, while Astra carries lower API prices.
The models are available to ChatGPT Work and Codex users with Plus, Pro, Enterprise and Edu subscriptions, as well as through the API. OpenAI highlighted improvements in coding, document preparation, presentations, spreadsheets and scientific research.
Benchmark results and API pricing
GPT-6.2 Astra scored 78.3% in Agents’ Last Exam, a test of an AI system’s ability to complete multi-step professional tasks. Its predecessor scored 73.1%, while Anthropic’s Claude Fable 5.2 scored 72.8%.
In the Artificial Analysis Intelligence Index, Claude Fable 5.2 led with 64.6 points, followed by GPT-6.2 Astra with 62.1 and Gemini 4 Carbon with 60. Claude Opus 5.5 scored 59.1, while GPT-6.2 Sol received 57.3 points, according to performance data published by OpenAI.
In FrontierCode 2.0, which evaluates changes to existing codebases, Sol scored 43.8%, narrowly trailing Astra’s 44.2%. The cost of completing tasks with Sol was about five times lower.
OpenAI reduced Astra’s API price for processing 1 million input tokens from $10 to $8 and its price for 1 million output tokens from $50 to $40. The price for cached input tokens fell from $1 to $0.40 per million tokens.
Sol’s corresponding API prices remained unchanged at $2 for input tokens, $10 for output tokens and $0.10 for cached input tokens per million.
OpenAI also disclosed an issue found during its safety evaluation. In two tests, the new models were more likely than their predecessor to conceal circumstances that prevented them from completing assigned tasks, including missing required files or a lack of search access.
Early testers criticize model performance
The models received mixed feedback from early testers despite their benchmark results. Every CEO Dan Shipper called the release disappointing compared with Anthropic’s Claude Fable 5.2.
Shipper said GPT-6.2 Astra trailed the previous version in writing quality, tended to be excessively brief, stopped coding work prematurely and struggled to produce varied designs.
Investor and former HyperWrite executive Matt Shumer declined to publish a full review of Astra and Sol because of unsatisfactory test results. He described the models as essentially unusable.
T3 Chat founder Theo Brown called GPT-6.2 Astra the worst model OpenAI has released. Wharton School professor Ethan Mollick offered a more measured assessment, saying the model was harder to steer than competing systems but could produce strong results on large-scale projects.
OpenAI’s enterprise customers reported more positive test results. ClickUp head of AI Jay Hack said token use fell by about 40% and the number of refinement cycles declined by 30%.
Source: Incrypted
