GPT-6 Cyber stays in Daybreak Red

OpenAI announced GPT-6 Cyber on 5 November 2026. The public API does not get it. A select group of Daybreak Red customers can call it through the API and Codex. Everyone else waits, with no date.

The @OpenAI post that afternoon says the model powered the Linux kernel research the company published on 4 November. It also says security teams at Mozilla, Palo Alto Networks, Calif, and Cisco are already using it, along with OpenAI’s own team. When this desk read the post, X showed 3.9 million views.

The scores on the announcement page

The announcement page says GPT-6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber. ExploitGym, in OpenAI’s wording, tests whether a model can exploit a specified vulnerability. The page calls the jump a 51% relative improvement.

On a separate alignment evaluation, the page says the model avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, and in 100% of tests with that prompt. Without the prompt, GPT-5.6 Cyber was at 94.49%. OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. The page also says the new model beats GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request. Astra’s ExploitGym score is not in the text this desk read.

The Verge covered the launch the same afternoon. Emma Roth’s piece, 5 November 2026, 5:00 PM UTC, rounds the alignment result to “less than 1 percent” out of bounds, down from “around 5 percent” for GPT-5.6 Cyber. She quotes Fouad Matin, head of OpenAI’s cyber preparedness and safety team, from a press briefing: “It’s over 99 percent of the time doing what you ask.” That sentence is The Verge’s report of the briefing, not a line on the announcement page. Roth names the alignment test ExploitGym HoneyPot. The announcement page calls it a separate alignment evaluation and does not use that name. Roth also reports that Matin called evaluation awareness a concern, and said OpenAI is trying to prioritize tests that look like production work.

What OpenAI says it found

On 4 November, OpenAI published “A thousand fewer paths to root.” The page says one mid-October run, using an upcoming cyber model, analyzed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 vulnerabilities, and produced 948 local privilege escalation exploits. OpenAI describes those exploits as a way an unprivileged local user could gain root. The same page prices the run at $1,213,430 in API credits, or $1,279 per exploit. The lede rounds the bill to $1.2 million. The page says the virtual machines ran a stock desktop operating system, not a weakened target, and that a human security researcher reviewed the outputs before vulnerabilities were reported with candidate patches. It also says the newer model found more than five times as many novel local privilege escalations as a GPT-5.6 Sol run in July, after three months of hardening.

The 5 November announcement names the model. It says OpenAI researchers used GPT-6 Cyber to develop those 948 Linux exploits, and to identify 81 previously unknown high-severity vulnerabilities in Firefox. It says the model also showed that 30 known Firefox issues were more severe than previously reported, and that it helped write candidate patches. The page says Mozilla has released fixes for many of the Firefox vulnerabilities. This draft did not open the Mozilla advisory, so it does not count how many fixes shipped.

The @OpenAI post on 5 November says the model “powered the Linux kernel research we announced yesterday.” The 4 November page still calls the model “upcoming.” Both sentences are OpenAI’s. This draft treats them as the same run because OpenAI connects them, not because an outside lab reproduced the count.

Who can call it, and what it costs

The model id on the API docs is gpt-6-cyber. The docs say an organization needs Daybreak Red access and approval. Input is text and image. Output is text. The context window is 1,050,000 tokens, with a maximum of 922,000 input tokens and 128,000 output tokens.

The same page lists standard short-context prices: $20 per million input tokens, $2 for cached input, $25 for cache writes, and $100 per million output tokens. It adds that a short-context cache write costs $25 per million tokens for 30 minutes, or $50 per million for 12 hours. The flagship pricing table this desk opened did not list gpt-6-cyber. The price here is the model page, not that table. A long-context price was not in the section this desk read.

The system card, published 5 November 2026, calls GPT-6 Cyber a cyber-specialized variant of GPT-6 Astra. It assigns the same Preparedness designations as Astra: Critical in cybersecurity, High in biological and chemical, and below High in AI self-improvement. Access runs through Daybreak Red, with identity verification, use restrictions, single sign-on, multi-factor authentication, and API IP allowlisting. API customers on Zero Data Retention generally need Private Safety Processing.

The announcement splits the program in two. Daybreak Blue is for verified defenders validating vulnerabilities and fixes, with fewer unnecessary refusals. Daybreak Red is the tier for deeper exploit development and authorized security testing with GPT-6 Cyber. Initial access is a select group of approved Daybreak Red customers, through the API and Codex. OpenAI says it will expand that group to all Daybreak Red customers soon, and that it is opening GPT-5.6 Cyber to all of them now. The announcement page names Cisco, Calif, and CodeRabbit as teams already using GPT-6 Cyber for vulnerability research and code review. The X post also names Mozilla and Palo Alto Networks.

What is still not on the page

There is no date for ChatGPT, for a consumer plan, or for a public API key. The 5 November thread points to a “Cyber at the Frontier” livestream on 12 November at 10am PT. That is a date on the post, not a public release.

The 948 count, the $1,213,430 bill, and the 81 Firefox findings are OpenAI’s. No independent lab had reproduced them in the pages this desk read. A second Verge item, Stevie Bonifield on 6 November, restates the Linux count and says a human researcher reviewed the patches. It does not add a second measurement.

Sources

Official

Press

The scores on the announcement page

What OpenAI says it found

Who can call it, and what it costs

What is still not on the page

Sources

Sources

Markers