Skip to content

OpenAI rates new GPT-6.2 Astra model at critical cybersecurity risk level

Category
News
Published
Reading time
2 min

OpenAI cut Astra's API price by 20% and rated the new model at a critical cybersecurity risk level, pairing it with a Control Room meant to catch risky actions early.

OpenAI released two new models, GPT-6.2 Astra and Sol, on Nov. 17. Astra is the first model OpenAI has shipped that crosses the company's high-risk line for machines improving themselves, and together with Sol it reached the critical line for cybersecurity risk under OpenAI's preparedness framework.

OpenAI said it deployed a Control Room system meant to halt suspicious model activity before it causes harm.

Astra's API price fell 20% from its predecessor, now costing $8 per million tokens of input and $40 per million tokens of output. Sol's rate held at $2 and $10. Both models are available through ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, as well as through the API.

Astra scored 62 on Artificial Analysis's Intelligence Index, placing second behind Anthropic's Claude Fable 5.2 at 65, according to Artificial Analysis. On the Agents' Last Exam benchmark, Astra beat Claude Fable 5.2 by 5.5 points while using about 5% of the cost.

Epoch AI found that on real-world tasks, Astra trails Claude Fable 5.2 and only edges out the previous GPT-6 Astra, and said independent testing showed no improvement in coding-agent performance. The model “can reason this way without showing their work,” the researchers said, complicating oversight.

Artificial Analysis measured high hallucination rates for both models: 59% for Astra and 55% for Sol.

Astra, which OpenAI built in part for legal work, ranked just 11th on the HLAB benchmark according to Vals AI, with Sol placing 21st.

OpenAI has not yet added Astra to its Daybreak program, which gives cybersecurity researchers expanded access to its models.