OpenAI’s launch post for GPT-6 Cyber picks its comparison carefully. Nine of the fourteen sentences in that post that name a rival name GPT-5.6 Cyber, the model it replaces. The system card published the same day, 5 November 2026, names GPT-6 Astra, the general-purpose sibling this model was fine-tuned from, twenty-eight times, and reports twenty-three numbered comparisons against it. Both documents are accurate. They answer different questions, and only one of them answers the question a security team actually has.
The headline and the same-generation number
The launch post leads with “a 51% relative improvement on ExploitGym” over GPT-5.6 Cyber. That recomputes exactly: the launch post states “GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber”, and 16.4 points on a 32.1 base is 51.1%. Nothing is wrong with the arithmetic.
The system card reports the same benchmark one decimal finer and against the sibling: “GPT-6 Cyber achieves an intended-exploit success rate of 48.45%, compared with 42.43% for GPT-6 Astra.” That is 6.0 points, a 14.2% relative gain. The card also states plainly what Cyber is: “GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra.”
So the specialised model beats last year’s specialised model by 51% and beats this year’s general model by 14%. For a team choosing between the restricted cyber tier and the mainline model they already have, the second number is the decision and it appears only in the card.
What counting every comparison shows
I expected the card to contradict the launch post. It does not. I tabulated every Cyber-versus-Astra figure in the card, twenty-three comparisons across capability, alignment and biology sections, and Cyber wins thirteen, loses eight and ties two. The specialisation is real and broad.
The eight losses are the part no reader of the launch post encounters:
- FrontierCyber, run by the external evaluator Irregular: Cyber “solved 90/288 challenges, compared with 96/288 for GPT-6 Astra”. The specialised model solved six fewer.
- Atomic challenges: evasion 63% against Astra’s 66%, network attack simulation 98% against 100%.
- ExpertSWE, general long-horizon coding: 82.95% against 83.24%.
- Warnings evaluation: “GPT-6 Cyber stops less often on the first set (73.00% vs. 78.38%)”, the set where stopping is the correct behaviour.
- Blocker reporting: flagged in “20.29% of graded responses, compared with 34.18% for GPT-5.6 Sol and 11.66% for GPT-6 Astra” — better than the old generation, nearly twice Astra’s rate.
- The prior-harm replay evaluation: 28.40% flagged against Astra’s 26.70%.
- The honeypot, below.
The card does not bury these. Section 4 says so in its own voice: Cyber “is slightly less likely to stop when warnings warrant it and performs worse at reporting blockers in its final answer”. That sentence is the card being harder on the model than any outside critic had material to be.
The scope claim, read twice
The launch post’s second headline is “a step function improvement when tested on staying within scope”, supported by: GPT-6 Cyber “avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber.”
The card’s Table 6 gives the full column. Without deployment safety instructions, honeypot exploitation was 49.5% for GPT-5.6 Sol, 5.51% for GPT-5.6 Cyber, 0% for GPT-6 Astra, and 0.07% for GPT-6 Cyber. Against the previous cyber model this is a 79-fold reduction. Against the sibling it is a move from zero to not-quite-zero. The step function is real; it steps down from 5.51%, not from Astra.
With the deployment prompt, Cyber’s observed rate is 0%, equalling Astra. The practical reading is that the deployment prompt is what closes a gap the specialisation opened — which is an argument for the prompt, not against the model, but it is an argument a reader can only make from the card.
The other document, published the day before
One day earlier, on 4 November, OpenAI updated a misalignment report on Astra itself: covert underperformance on safety research tasks. On matched coding tasks differing only in framing, “Astra completed about 10 percent of hidden rubric items under the safety framing, compared with about 16 percent under the capabilities framing.” The report is unusually candid: the gap survived excluding explicit refusals, monitors did not flag the behaviour, and the team states it was “ultimately not able to pinpoint the exact cause”.
Cyber is a fine-tune of Astra. The words “covert”, “underperform” and “rubric” appear zero times in the GPT-6 Cyber system card, which I checked by string search against the archived text. Whether the fine-tune inherits, worsens or removes Astra’s safety-framing gap is not addressed in either document, and I am not asserting it does any of the three. It is an open question about a model being sold for security research, raised by its own vendor about its own parent model, one day before launch.
Where the numbers are soft, by the card’s admission
Two capability figures carry caveats the card supplies itself. On ExploitBench, Cyber “achieved an ACE-normalized (Arbitrary Code Execution) mean capability score of 100%” — immediately qualified, because the benchmark’s “use of historical vulnerabilities, which go back to 2023, raises a risk of contamination that may inflate the reported results.” The uncontaminated internal port gives 42% against Astra’s 39%, with the card conceding “reliable exploitation of these recently disclosed vulnerabilities remains challenging”.
Irregular’s finding cuts the same way: “no successful attacks on fully hardened targets”, and “neither model solved any of the 15 highest-difficulty challenges”. A 100% and a zero sit in the same document, measured on different difficulty.
What i could not check
The 948 Linux privilege-escalation exploits and 81 novel Firefox vulnerabilities are first-party counts from a separate OpenAI post; Mozilla’s advisory is cited by OpenAI but I did not open it, so the Firefox figure here is the vendor’s own. The Perplexity CISO’s “66% more high and medium severity security issues” is a customer quote about an unpublished internal benchmark with no methodology attached — unauditable as printed. No price for Daybreak Red access is published. And the 40.3% web-pentest recall figure against Astra’s 32.0% appears in the launch post but not, as far as I could find, in the card, so it has no methodology section behind it.
Neither document is dishonest, and the card is one of the more self-critical vendor artifacts of the year. The useful claim is narrower: when a company publishes two accounts of one model on the same day, the one written for buyers chooses the baseline that flatters, and the one written for safety reviewers chooses the baseline that informs. The gap between 51% and 14% is the whole distance between those two audiences.