OpenAI ships GPT-6 Cyber at twice Anthropic's price and five times Google's
OpenAI released GPT-6 Cyber on Nov. 5 at $20 input and $100 output for each million.
The decision it puts on your desk
Test one finding from your current tool against this model by Nov. 19, then decide whether $20 input beats a $10 rival.
OpenAI shipped GPT-6 Cyber on Nov. 5, and it sells at twice Anthropic's rate for each million input.
Only approved customers can buy it at first, through a program OpenAI calls Daybreak Red. OpenAI rates the model Critical for security work, a step up from High under its own rules.

Pricing above the field
OpenAI built the model on its Generative Pre-trained Transformer (GPT) line, then trained it further on security work. It lists at $20 input and $100 output for each million. Anthropic's Claude Mythos 5.1 sells at $10 and $50, while Google's Gemini 4 Argon Cyber lists at $4 and $20.
Rival labs price the same kind of work at a quarter of that rate. Buyers at five times Google's rate are paying for deeper results on real work. OpenAI published no measure of cost per useful finding, only token rates.
Exploit benchmark results
On ExploitGym, which scores a model on a given weakness in a remote target, GPT-6 Cyber reached 48.5 percent. GPT-5.6 Cyber scored 32.1 percent on that test last year. A second test of web applications found 40.3 percent against 32.0 for GPT-6 Astra.
A new evaluation called RemoteExploitBench put known weaknesses in GitLab, PostgreSQL and Jenkins. GPT-6 Cyber completed 95 percent of those tasks. GPT-6 Astra under a lighter access tier refused every one of them.
Findings on Linux and Firefox
OpenAI reported that researchers using the model reproduced 1,724 weaknesses in the Linux kernel. They built working attacks for 948 of them, against 166 from GPT-5.6 Sol on the same setup. OpenAI sent 304 critical findings to Linux maintainers with candidate patches.
"Firefox entered 2026 as one of the world's most hardened and scrutinized codebases," said Brian Grinstead, a distinguished engineer on Firefox.
Mozilla began landing fixes after the reports, and it has rated 81 of them high severity. OpenAI also argues that 30 known issues deserved a higher rating.
"GPT-6 Cyber is the strongest model we've tested for security review," said Harjot Gill, chief executive of CodeRabbit.
Scope and access controls
OpenAI's release leans hard on staying inside an assigned task. On a test built after the Hugging Face incident, it avoided side targets in 99.93 percent of runs. GPT-5.6 Cyber managed 94.49 percent on that test without special prompts.
OpenAI requires approved network addresses for program access by default. It also applies automated checks for misuse and for stolen credentials. OpenAI added a separate group for teams that run critical plants.
Google and Anthropic sell rival security models at far lower prices. Both must now decide whether to hold rates or match a rival that claims deeper results. That comparison stays open until outside teams publish cost per useful finding across all three.
Sources
← Previous
Morgan Stanley put $15 trillion on AI. The fund managers in the same room want the returns first.
Next →
OpenAI is buying Pierre Computer. Microsoft owns GitHub and nearly a quarter of OpenAI.
The Morning Recap
We sift the day. You make the calls.
The decision on every story, in your inbox before your first call.