Key Findings
GPT-6 Cyber demonstrates a marked increase in exploit research capabilities, achieving a 48.5% score on ExploitGym compared to 32.1% for GPT-5.6 Cyber. In tests on web applications without source code access, the model identified 40.3% of vulnerabilities, surpassing the 32% found by GPT-6 Astra. Furthermore, GPT-6 Cyber completed 95% of tasks on RemoteExploitBench, a benchmark involving applications such as Jenkins, PostgreSQL, and GitLab.
Technical Details
A variant of GPT-6 Astra, GPT-6 Cyber underwent additional training for security investigations. OpenAI utilized reinforcement learning to reward the reproduction of crashes, the construction of exploits, and the accurate description of results. Under OpenAI’s Preparedness Framework, the model’s cybersecurity capability is rated as “Critical,” an increase from the “High” rating assigned to GPT-5.6 Cyber.
Real-world application data shows that researchers using the model reproduced 1,724 Linux kernel vulnerabilities and developed 948 working exploits. In the case of Firefox, the model identified 81 previously unknown high-severity vulnerabilities and provided evidence that 30 known issues required higher severity assessments.
Background & Context
Access to GPT-6 Cyber is managed through the gated “Daybreak Red” program, which requires approval and documentation of specific defensive workflows. As of last month, OpenAI has enrolled over 2,000 entities across its cybersecurity programs, including the less restrictive “Daybreak Blue” program used for GPT-6 Astra.
Significance & Outlook
The model carries a premium price point: $20 per million input tokens and $100 per million output tokens. This represents a 60% increase in input costs and a 33% increase in output costs compared to GPT-5.6 Cyber. When compared to rivals, GPT-6 Cyber’s base rates are twice those of Anthropic’s Claude Mythos 5.1 and five times those of Google’s Gemini 4 Argon Cyber Permissive. While CodeRabbit reports finding 66% more medium- and high-severity issues on internal benchmarks using GPT-6 Cyber, the high cost per token remains a primary differentiator against competing cyber models.
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
