GPT-6.2 Astra: The first publicly available model for automated AI development
After the planned October release of GPT-6.1 Astra was canceled due to security concerns, OpenAI has now officially announced the next generation of GPT-6.2 Astra and GPT-6.2 Sol. Both are available to Plus, Pro, Enterprise, and Edu subscribers via ChatGPT Work and Codex. The most demanding Pro users also get access to the faster Ultrafast mode.
The two new models are designed for different types of tasks. Sol remains the default choice for everyday work and maintains its current price. The more powerful Astra is designed for more demanding professional tasks, such as preparing financial models, legal documents and presentations, and programming. Developers will pay approximately €7.29 for one million input tokens on the Astra model, and around €36.46 for one million output tokens. The lower costs are especially important when using autonomous agents, which can generate huge amounts of data during long-term task execution.
OpenAI states that Astra scores 5.2 points better than its predecessor on the Agents' Last Exam test and 5.5 points better than the competing model Claude Fable 5.2. It also reportedly achieved the best result in the world on the Terminal-Bench-Science test. Despite the promising results, it should be noted that these are manufacturer's statements. The independent Artificial Analysis Index ranks Astra in second place, right behind Anthropic's Claude Fable 5.2. Astra does not therefore lead in all comparative tests.
Of particular note are the security ratings. Both new models have been rated critical for cybersecurity risk and high for biological and chemical capabilities in OpenAI’s Security Readiness System. As a result of the critical cybersecurity rating, OpenAI is implementing additional safeguards and controls over the use of the models. These could also limit some of the legitimate activities of security researchers.
The main concern is the ability of new models to accelerate their own development. The previous generation GPT-6 Astra has already shown the ability to change the weights of another model during testing, and also a reduced willingness to participate in some security research. This raises an important question: will developers still be able to thoroughly test the security of increasingly powerful models before offering them to the general public?



















