OpenAI Named the Model Behind the 948 Linux Root Exploits. GPT-6 Cyber Costs $20 In and $100 Out, and Only Daybreak Red Customers Get It.
On November 5 OpenAI introduced GPT-6 Cyber, a cyber-specialized variant of GPT-6 Astra and the model behind the previous day's Linux kernel run. It is rated Critical for cybersecurity under the Preparedness Framework, up from High for GPT-5.6 Cyber, and is available only to approved Daybreak Red customers through the API and Codex, at $20 per million input tokens, $100 output, and $2 cached, with a 1,050,000-token context. OpenAI's own evaluations put ExploitGym intended-exploit success at 48.5% against 32.1% for GPT-5.6 Cyber, and out-of-scope exploitation on its honeypot at 0.07% without the safety prompt. Secondary price checks put it at roughly twice Claude Mythos 5.1 and five times Gemini 4 Argon Cyber Permissive on list. We read the launch post, the system card, and the model docs.
By T.W. Ghost
The "upcoming cyber model" in yesterday's Linux kernel post has a name. OpenAI introduced GPT-6 Cyber today, November 5, 2026: a cyber-specialized variant of GPT-6 Astra, model id gpt-6-cyber, rated Critical for cybersecurity under the Preparedness Framework, with High for biological and chemical and below High for AI self-improvement, the same profile as Astra. GPT-5.6 Cyber was rated High. Access is approved Daybreak Red customers only, through the API and Codex, expanding to all of Daybreak Red; GPT-5.6 Cyber opens to all of Daybreak Red at the same time, and the Daybreak Blue alpha for Astra is expanding.
The price is on the model page: $20 per million input tokens, $100 output, $2 cached input, with cache writes at $25 for a 30-minute window or $50 for 12 hours. Context is 1,050,000 tokens, with 922,000 of input and 128,000 of output, text and image in, text out. By VentureBeat's check of the docs, GPT-5.6 Cyber listed at $12.50, $75, and $1.25, so the new model is 60% more on input and about a third more on output; the same check puts Claude Mythos 5.1 at $10 and $50, which matches Anthropic's page, Gemini 4 Argon Cyber Permissive at $4 and $20, and Gemini 3.8 Flash Cyber at $1.50 and $7.50. Every capability number below is OpenAI's. We read the launch post, the system card, the model docs, and the VentureBeat price check. Here is what held up.
What You Are Applying For
| GPT-6 Cyber (November 5, 2026) | |
|---|---|
| Base | Cyber-specialized variant of GPT-6 Astra |
| Access | Approved Daybreak Red customers; API and Codex; expanding to all Daybreak Red |
| Price | $20 in / $100 out / $2 cached per 1M; cache writes $25 (30 min) or $50 (12 h) per 1M |
| Context | 1,050,000 total; 922,000 max input; 128,000 max output; text and image in |
| Preparedness | Critical cybersecurity; High bio and chem; below High self-improvement (same as Astra) |
| GPT-5.6 Cyber list | $12.50 / $75 / $1.25 (VentureBeat check of OpenAI docs) |
| Rival list prices | Mythos 5.1 $10 / $50 / $0.25; Gemini 4 Argon Cyber Permissive $4 / $20 / $0.20; Gemini 3.8 Flash Cyber $1.50 / $7.50 / $0.15 (VentureBeat; Mythos confirmed on Anthropic's page) |
| Controls | IP allowlisting default for Daybreak Red API; zero-data-retention customers generally must use Private Safety Processing with 30-day encrypted customer-side retention |
| Programs | Dragos partnership for operational-technology safeguards; new Daybreak Red for Critical Infrastructure group; $1 billion Daybreak for Frontline Defenders grant still cited |
Gated and expensive are the same decision. Twice Mythos and five times Argon Cyber Permissive on list, behind an application, is OpenAI choosing who runs this and how much they think about it per token. Token price is not cost per finding; nobody has published that for any of the three.
Critical is shared with Astra. The designation describes the threshold a model clears, not a head-to-head. Irregular's updated FrontierCyber suite, cited in the system card, has GPT-6 Cyber at 90 of 288 and Astra at 96 of 288, which is the one place in the materials where the specialized model does not win.
The Benchmarks, All of Them OpenAI's
| Claim | GPT-6 Cyber | Comparator | Who |
|---|---|---|---|
| ExploitGym intended-exploit success | 48.5% | GPT-5.6 Cyber 32.1% | Vendor |
| Web pentest recall | 40.3% | Astra 32.0% | Vendor |
| RemoteExploitBench completion | 95% | Astra with Daybreak Blue refused 100% | Vendor |
| ExploitGym Honeypot, out-of-scope exploitation, no safety prompt | 0.07% (99.93% stayed in bounds) | GPT-5.6 Cyber 5.51% | Vendor (system card) |
| Same, with the deployment safety prompt | 0% | Vendor | |
| Linux kernel run | 948 unique LPEs; 1,724 reproduced; $1.2M; 304 critical findings disclosed with candidate patches | Vendor; the Nov 4 post | |
| Firefox | 81 novel high-severity bugs by Mozilla's classification plus 30 known issues re-rated; Mozilla landing fixes | Vendor discovery; Mozilla severity | |
| CodeRabbit | 66% more high and medium findings than its production model | Vendor quote |
Three honest observations:
- The alignment pitch is scope, and the test is OpenAI's. Staying in bounds 99.93% of the time without a safety prompt is the number OpenAI leads with, and the motivation it cites is the Hugging Face incident. It is measured on OpenAI's own honeypot. It is not a production guarantee, and OpenAI does not claim it is.
- The 948 is yesterday's number. The launch post repeats the Linux run; it does not add exploits. The new disclosure here is Firefox, where the severity ratings are Mozilla's and the discoveries are OpenAI's.
- The customer quotes are quotes. Mozilla, Calif, Cognition, Palo Alto, Cisco, SentinelOne, CrowdStrike, Perplexity, and CodeRabbit are named. One gives a number. None of it is an independent evaluation.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Announced Nov 5; variant of Astra; Daybreak Red only; API and Codex | Vendor (launch post; system card) |
| $20 / $100 / $2; cache writes $25 and $50; context 1,050,000 | Verified (OpenAI model docs) |
| Critical cyber; High bio and chem; below High self-improvement | Vendor (system card) |
| ExploitGym, RemoteExploitBench, honeypot, pentest recall figures | Vendor |
| Firefox 81 novel high-severity | Vendor discovery; Mozilla classification |
| GPT-5.6 Cyber and rival list prices | Secondary (VentureBeat); Mythos $10 / $50 confirmed on Anthropic's page |
| IP allowlisting; Private Safety Processing for ZDR | Vendor (system card) |
| Independent replication of any score | None |
Who Should Use It
If you are already in Daybreak Red, check whether GPT-6 Cyber is provisioned on your account, then re-price the workloads you run on it against Mythos and Argon Cyber, because the gap on list is large and the gap in findings per dollar is unmeasured.
If you are applying, plan for IP allowlisting, SSO and MFA, and Private Safety Processing if you are a zero-data-retention shop; those are conditions of the gate, not options.
The frontier is elsewhere: three labs now sell gated cyber models, and the week's most concrete security consequence for everyone else is still the patch backlog from the Linux run. Our model comparison has the current standings on the models you do not need an application to use.
Sources
- OpenAI: introducing GPT-6 Cyber
- OpenAI deployment safety: GPT-6 Cyber system card
- OpenAI API docs: gpt-6-cyber
- OpenAI: a thousand fewer paths to root
- VentureBeat: OpenAI launches GPT-6 Cyber with stronger exploit research and a higher API price than Anthropic and Google (price check)
- Anthropic: Claude Mythos
*Not sure which lab's cyber gate your security program can get through, or afford? Take the free 2-minute quiz and get matched. Then read about Anthropic's three-tier version of the same door: the Cyber Verification Program.*