OpenAI released GPT-6 Cyber on Nov. 5, 2026. The model hunts software flaws and writes working exploits for teams running authorized security tests1.
Few professionals can touch it. OpenAI locked GPT-6 Cyber behind Daybreak Red, its approval program for advanced cybersecurity work. The company priced the model at $20 for every million words it reads and $100 for every million it writes. That price doubles what OpenAI charges for Astra, its own general-purpose model12.
The Price of the Gate
Daybreak Red requires an application. An organization applies, submits its defensive workflow, and clears a security review before OpenAI opens the door. The new model then needs a second, separate approval on top of that2.
Individuals face their own checklist. OpenAI requires single sign-on, multi-factor authentication, managed devices, and IP addresses on an approved list. A solo researcher needs Advanced Account Security. That is OpenAI's term for a login protected by a physical hardware key2.
Rivals charge less for a similar seat at the table. Anthropic's Claude Mythos 5.1 costs $10 for a million input tokens and $50 for a million output tokens. That matches Claude Fable 5.1, its generally available twin. Anthropic reserves Mythos 5.1 for organizations cleared by its Cyber Verification Program5.
Google runs its own gated tier too. Gemini 4 Argon's Cyber Permissive option has billed $4 and $20 through Fairwind, Google's approval program. VentureBeat reported those figures1.
GPT-6 Cyber Charges Double What Its Rivals Bill
Output price per million tokens for three gated security models, checked Nov. 5, 2026
Each column is one model's output price; the magenta column is GPT-6 Cyber, and the gray columns are the two rivals it costs twice as much as.
Each model also bills a separate input price: GPT-6 Cyber's $20 is double the $10 that GPT-6 Astra and Claude Mythos 5.1 both bill.
Sources: VentureBeat [1]; Claude Platform Docs [5]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| GPT-6 Cyber | $100 |
| GPT-6 Astra | $50 |
| Claude Mythos 5.1 | $50 |
The new model costs more than either rival bills for the same lane of work. A security team that clears every checkpoint still pays double, whichever one it picks15.
What the Gate Buys
Money and paperwork buy real capability. GPT-6 Cyber scored 48.5 percent on ExploitGym. That is OpenAI's own test of whether a model can turn a known flaw into a working exploit. The score is up from 32.1 percent for GPT-5.6 Cyber, the model it replaces1.
GPT-6 Cyber Still Misses Half the Exploits It Chases
Two of OpenAI's own benchmarks, percent, GPT-5.6 Cyber against GPT-6 Cyber
Each row is one test; the gray dot is the older model's score, the magenta dot is GPT-6 Cyber's score, and the rod marks the gap it closed.
OpenAI ran both tests itself, and an outside lab has yet to confirm either score.
Sources: VentureBeat [1]; SecurityWeek [3]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | GPT-5.6 Cyber | GPT-6 Cyber |
|---|---|---|
| Solved ExploitGym tasks | 32.1% | 48.5% |
| Stayed inside assigned scope | 94.49% | 99.93% |
The new model already found real bugs. Researchers pointed GPT-6 Cyber at Mozilla's Firefox browser. They came back with 81 new flaws Mozilla rated high severity, each backed by a reproducible test. Mozilla has started shipping fixes4.
Linux turned up more. The model reproduced 1,724 known kernel flaws and built working exploits for 948 of them. Of those, 588 were previously unknown. OpenAI sent 304 of the worst findings to the kernel's maintainers, candidate patches attached4.
Dragos, a company that defends power grids and factories, ran the model against a mock power plant. The model found two zero-day flaws in the software managing the plant's protective relays. Those relays are switches that cut power before a fault spreads4.
The Case for Paying Up
CodeRabbit tested the new model before OpenAI's price tag scared anyone off. The company dropped the new model into its own security scan pipeline and ran it against known bugs.
CodeRabbit@coderabbitaiWe put GPT-6 Cyber inside CodeRabbit’s security scan pipeline.
22 of 33 known vulnerabilities caught vs 19 for our production ensemble, and 66% more high and medium severity findings.
Full eval below. https://t.co/OyfVJ3JaDw pic.twitter.com/x92iR9NRp5
The result impressed CodeRabbit's chief executive. "On our internal benchmark, it found 66% more high and medium severity issues than our production model," Harjot Gill said2. CodeRabbit's test caught 22 of 33 known flaws with GPT-6 Cyber. Its production tool caught only 19 of the same flaws.
That gap is the strongest argument for paying OpenAI's price. A tool that catches a third more real bugs earns its seat, even a costly one that takes weeks to clear.
The Catch
Dragos answers Gill's case with its own warning. Rob T. Lee is the company's chief AI officer.
He watched the model work and drew a harder lesson than OpenAI's launch post did. "The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack," Lee said3.
Lee's point cuts both ways. The same model helps Mozilla and CodeRabbit catch bugs faster. It also lowers the bar for whoever gets past Daybreak Red's checks, or builds something similar outside them.
OpenAI's own numbers carry a second catch. The model still fails more than half of its ExploitGym tasks, the same test behind its 48.5 percent score1.
Its scope discipline improved too. GPT-6 Cyber stayed inside its assigned boundary 99.93 percent of the time, up from 94.49 percent for the model before it3. That still leaves room for a mistake inside a system the client's contract excluded.
The company itself admits the stakes. OpenAI rated GPT-6 Cyber "Critical" under its own safety framework. That is a step above the "High" rating it gave GPT-5.6 Cyber3. A tool rated that dangerous, that costly, and that locked down still loses more attempts than it wins.
What to do
Price the whole seat beyond the API call alone. Weigh the weeks a Daybreak Red review takes against the dollars the model bills. Then set that total beside Claude Mythos 5.1's review and Gemini 4 Argon's Fairwind process.
Ask a vendor for its own benchmark before buying the newest model on the shelf. CodeRabbit ran it against known bugs before trusting Gill's praise. Any buyer can ask a prospective tool to clear the same bar against a known test set.
Treat a 48.5 percent score as a floor, short of a verdict. Keep a human reviewer on every finding a model this new and this costly produces. Mozilla and the Linux kernel's maintainers already apply that same discipline to its reports.
By the numbers
- $100: GPT-6 Cyber's price for every million output tokens, double GPT-6 Astra and Claude Mythos 5.11.
- 48.5 percent: GPT-6 Cyber's score on OpenAI's ExploitGym test, up from 32.1 percent for the model it replaces1.
- 81: new high-severity Firefox flaws GPT-6 Cyber surfaced, each with its own reproducible test4.
- 948: working Linux kernel exploits GPT-6 Cyber built, 588 of them previously unlogged4.
- 99.93 percent: how often GPT-6 Cyber stayed inside its assigned scope, up from 94.49 percent3.
- 2: zero-day flaws GPT-6 Cyber found in the software managing a mock power plant's protective relays4.
- 66 percent: more severe bugs GPT-6 Cyber caught than CodeRabbit's own production security tool2.
What to watch
Watch OpenAI's promised rollout to every Daybreak Red customer. A wider release would test Lee's warning about a dropping skill floor against a wider circle of defenders carrying the same tool23.
Mozilla's and the Linux kernel's patch logs matter too. Fixes that trace back to the model's reports, arriving on a steady schedule, would settle the question. A catch rate this high should hold up past its first month4.
Sources
- Carl Franzen, "OpenAI launches GPT-6 Cyber with stronger exploit research and higher API price than Anthropic and Google," VentureBeat, Nov. 5, 2026, https://venturebeat.com/technology/openai-launches-gpt-6-cyber-with-stronger-exploit-research-and-higher-api-price-than-anthropic-and-google
- Rama Adi, "GPT-6 Cyber: what it does, who can get it, and the catch," eesel, Nov. 6, 2026, https://www.eesel.ai/blog/gpt-6-cyber
- Eduard Kovacs, "OpenAI Says New GPT-6 Cyber Model Hunts Flaws Without Crossing the Line," SecurityWeek, Nov. 6, 2026, https://www.securityweek.com/openai-says-new-gpt-6-cyber-model-hunts-flaws-without-crossing-the-line/
- Pierluigi Paganini, "OpenAI Rates GPT-6 Cyber Critical After Major Vulnerability Discoveries," SecurityAffairs, Nov. 6, 2026, https://securityaffairs.com/202546/ai/openai-rates-gpt-6-cyber-critical-after-major-vulnerability-discoveries.html
- Anthropic, "Claude Mythos 5.1," Claude Platform Docs, Sept. 1, 2026, https://platform.claude.com/docs/en/models/mythos-5-1/overview
