OpenAI's new security AI costs twice Anthropic's and finds software flaws for about $1,300 each

OpenAI on Thursday released GPT-6 Cyber, an AI model built to find and exploit security flaws in software, to a short list of approved customers. Its list price is higher than rival security models from Anthropic and Google. It is also the clearest sign yet that finding security holes is becoming a priced, repeatable service.

The headline result came in research OpenAI published shortly before the launch. OpenAI said the model built working "root" attacks, which give a low-level user full control of a machine, for 948 distinct flaws in Linux. Linux runs most of the world's servers. The run cost $1.21 million in usage fees, or about $1,279 per flaw.

Every one of those figures comes from OpenAI. No outside lab has repeated the work, and no shared, cross-vendor test exists to compare security models. Buyers are being asked to take the numbers on trust, for now.

A premium price against Anthropic and Google

GPT-6 Cyber lists at $20 per million tokens of input and $100 per million of output, according to a VentureBeat price comparison. Tokens are the small chunks of text that AI models bill by. That is twice the rate of Anthropic's Claude Mythos 5.1 and five times Google's top security model, Gemini 4 Argon Cyber Permissive.

It is also a step up from OpenAI's own GPT-5.6 Cyber, which charges $12.50 and $75. OpenAI is betting buyers will pay for results rather than shop on rate cards. VentureBeat notes that token prices are not the cost of a finished investigation. Tooling, isolated test machines and human review all add to the bill.

OpenAI's pitch is that the work is still cheap. It compared the Linux haul with historical prizes of $20,000 to $70,000 per exploit at hacking contests such as Pwn2Own. On that basis it values the 948 flaws at $18.9 million to $66.3 million, or 16 to 55 times the cost of the run. That is OpenAI's own framing, and contest prizes are not what a company would actually pay.

The trend matters more than any single valuation. A July run with an earlier OpenAI model found 166 such flaws for $513,690, about $3,100 each. The October run cut the cost per flaw by roughly 60%, even though three months of fixes had hardened the code in between.

Early customer feedback is positive but thin. Harjot Gill, chief executive of code-review company CodeRabbit, called it "the strongest model we've tested for security review." CodeRabbit said it found 66% more medium- and high-severity issues on its internal test. Mozilla has begun incorporating fixes into Firefox for 81 previously unknown flaws that OpenAI says the model found and Mozilla rated high severity.

Gated access and a "Critical" risk label

Access runs through Daybreak Red, OpenAI's vetted program for security teams that test systems by attacking them. Applicants submit details of their defensive work for review. API access generally requires pre-approved network addresses. OpenAI says more than 2,000 organizations had enrolled as of last month, but it has not named them. It plans to open GPT-6 Cyber to all Daybreak Red members over the coming weeks.

OpenAI describes GPT-6 Cyber as a version of its general-purpose GPT-6 Astra model with extra security training. It rates the model "Critical" for cybersecurity risk under its internal safety framework, the same rating Astra already carries and one step above GPT-5.6 Cyber. It says the new model stayed within its assigned targets in 99.93% of tests without extra safety instructions, against 94.49% for its predecessor. Chris Hughes of Noma Security, quoted by CSO Online, said even a tiny miss rate could mean unauthorized actions at scale.

OpenAI also disclosed a flaw in its own process. In an early run, the system claimed 200 working exploits. Human reviewers found that 20 of them reached full control through a different flaw than the one assigned. OpenAI added an independent check, and its researchers now review every exploit and fix before outside maintainers see it.

Finding flaws is getting cheaper than fixing them

For corporate security chiefs, the bottleneck is moving. Within 10 days of the October scan starting, 61 of the Linux flaws had been fixed in the main code and 43 more had fixes in testing, according to Help Net Security's account of OpenAI's report. Hundreds more were still being reported.

Hughes pointed to OpenAI's own "Patch the Planet" tracker, which he said shows a fix rate below 25%. Faster discovery, he warned, could widen the gap between bugs found and bugs fixed. Monthly disclosures of software flaws rose from 5,045 in January to 10,740 in August, The Hacker News reported, citing Google threat-intelligence data.

Detection may get harder too. OpenAI says the model used at least 10 attack techniques that outside experts had not seen before. Rob T. Lee of security firm Dragos, an OpenAI partner, suggested in CSO Online that could mean AI is creating new ways in, not just finding old ones faster.

OpenAI's advice to companies is blunt. An unpatched system where an attacker holds even a low-level account should be treated as "functionally equivalent to root access" for a well-funded attacker. That raises the cost of slow patching for every business that runs Linux.

Anthropic, Google and OpenAI now all sell gated security models, at list prices that differ by a factor of five. Buyers have no neutral way to compare them. Until one exists, each vendor's own exploit counts will keep setting the terms of the sale.