On November 18 JST, OpenAI announced new AI models, "GPT-6.2 Astra" and "GPT-6.2 Sol." They have begun offering them through ChatGPT Work, Codex, and APIs for developers. For ChatGPT Work and Codex, Plus, Pro, Enterprise, and Edu plans are eligible.

Astra, the flagship model, has enhanced its ability to create slides, spreadsheet files, and other documents according to existing templates. According to OpenAI, the unnaturalness often associated with AI-generated content has been reduced, allowing it to produce deliverables that adhere to specified templates and industry conventions, making them readily usable in professional settings. It also supports tasks such as operating design software to model machine parts.

According to the company's test results, in "Agents’ Last Exam," which measures multi-step professional tasks, GPT-6.2 Astra outperformed GPT-6 Astra by 5.2 points, and the cost per task was less than one-third. In the same evaluation, it also surpassed Anthropic's Claude Fable 5.2 by 5.5 points.

Meanwhile, Sol, the general-purpose model, is designed for daily tasks that involve continuous multi-step processes, such as document creation, research, and file editing.

In "FrontierCode 2.0," which measures whether changes made to actual code meet adoption criteria, when comparing both models at their maximum inference settings, Sol achieved a score difference of less than 0.5 points from Astra at approximately one-fifth of the cost.

Standard API pricing per 1 million tokens for Astra is $8 for input, $0.40 for cached input, and $40 for output. Compared to GPT-6 Astra, input/output is 20% cheaper, and cached input is 60% cheaper.

Sol's pricing remains unchanged from GPT-6.1 Sol, at $2 for input, $0.10 for reused input, and $10 for output.

With the increase in capabilities, OpenAI has also strengthened restrictions on requests related to high-risk cybersecurity misuse. Furthermore, Astra is stated to be the first publicly released model to achieve a "High" rating in the company's evaluation for self-improvement capabilities, which accelerate AI research.

Made with AI