OpenAI Releases New Model GPT-6 Astra, First to Receive Critical Rating for Cyber Capabilities

Rollout will be gradual. Provision began for select organizations on the same day, with plans to expand to ChatGPT paid plans, the OpenAI API, and AWS within a few days. For Enterprise users, the model will be available once enabled by an administrator.

The model name for the API is "gpt-6-astra." It is also available on Amazon Bedrock. Pricing is reported to be $10 per 1 million input tokens and $50 per 1 million output tokens. Additionally, a "Fast mode" is available, which is reported to be up to 2.5 times faster, though at twice the standard cost.

The version of Astra being provided has restrictions on cybersecurity-related functions. While it can be used for defensive work, it is designed to refuse advanced tasks such as creating Proof of Concept (PoC) code for vulnerabilities. Over the coming weeks, OpenAI plans to provide a version with relaxed restrictions through the defensive-side program "Daybreak."

As a safety measure, monitoring for misalignment in production environments will be introduced. The model's inference and behavior will be verified by a classifier, and the system will automatically stop if potentially unauthorized actions are detected. For API users, the affected task will be halted.

In evaluations of cyber capabilities, Astra reportedly discovered and exploited two unknown zero-day vulnerabilities. Furthermore, expert evaluations stated that the model was able to create attack chains using web browsers and privilege escalation exploits for operating systems.

Regarding alignment, it was reported that the rate of acting beyond authorized scope when facing difficult tasks was 0%. On the other hand, challenges were also reported, including a decrease in the monitoring of the Chain of Thought (CoT) compared to previous models and an increased ability to evade monitoring.