GPT-6 Cyber

GPT-6 Cyber, which OpenAI released on November 5, 2026, costs twice as much as GPT-6 Astra at $20 / $100 per million tokens and beats it on most of OpenAI's hacking tests by 2 to 7 points, but only vetted security teams can use it.

The benchmarks

All numbers come from OpenAI's system card, at the highest reasoning budget OpenAI tested. OpenAI only compared GPT-6 Cyber against its own models, mostly GPT-6 Astra, the general model it was trained from. The Irregular rows are an outside lab's tests, commissioned by OpenAI and published in the same card. On FrontierCyber, GPT-6 Cyber solved 90 of 288 challenges to Astra's 96, and neither model beat a fully hardened target. The safety number OpenAI leads with: when a side system it wasn't cleared to touch sat next to the target, GPT-6 Cyber attacked it in 0.07% of runs, against 5.51% for GPT-5.6 Cyber, 49.5% for GPT-5.6 Sol and 0% for Astra.

OpenAI's numbers

• ExploitGym (exploiting a live target, max): 48.45%. +6.02 over GPT-6 Astra at max (42.43%).

• ExploitGym Hardened (exploits against stronger defenses, max): 36.94%. +7.25 over GPT-6 Astra at max (29.69%).

• SRE-Bench (reverse engineering programs, max): 92.84%. +4.86 over GPT-6 Astra at max (87.98%).

• SEC-Bench Pro (finding bugs in JavaScript engines, max): 87.02%. +1.64 over GPT-6 Astra at max (85.38%).

• ExploitBench Internal Port (exploits for fresh V8 and Chrome bugs, max): 42%. +3 over GPT-6 Astra at max (39%).

• ExploitBench (turning known bugs into exploits, max): 100%. Ties GPT-6 Astra at max (100%).

• ExpertSWE (long software engineering tasks, max): 82.95%. −0.29 under GPT-6 Astra at max (83.24%).

Independent numbers

Artificial Analysis, Vals and LMArena haven't tested it, since access is gated.

• Irregular CyScenarioBench (long multi-step attacks): 61%. +7 over GPT-6 Astra (54%).

• Irregular FrontierCyber (bugs in real software and hardware): 31.3%. −2.1 under GPT-6 Astra (33.3%).

First impressions

GPT-6 Cyber is a few hours old and gated, so almost nobody outside OpenAI's partners has touched it. Most of the reaction so far is about the Linux kernel research OpenAI posted on November 4, which OpenAI says this model powered.

The positives

zahlman on Hacker News pulled the headline number from OpenAI's kernel post.

We've developed an autonomous vulnerability discovery, exploitation, and patching pipeline. In one run with our cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in Linux at an API cost of $1.2 million. (comment)

Their follow-up was the hopeful read: imagine other open-source projects getting that kind of budget for security.

treebeard901 on Hacker News flagged the line in OpenAI's post they think matters most for everyone else.

Large, complex software projects likely require multiple rounds of scanning and fixing vulnerabilities to "bottom out" latent bugs. Ideally this process is repeated with each subsequent generation of frontier models. (comment)

If that holds, every new model means another round of patches for anyone who ships software.

The negatives

bananaquant on Hacker News pushed back on the counts.

From his talk, AIs can't reliably deduplicate or count vulnerabilities, dash for easy and meaningless escape hatches like malicious filesystems, and report the same things as found by everyone else. (comment)

"His talk" is Linux kernel maintainer Greg Kroah-Hartman's, about earlier AI bug reports. tzury made a similar point: one flaw can show up as many "exploits." OpenAI hasn't published the full list, so nobody outside can check yet.

evan_a_a on Hacker News didn't buy the reason for holding back details.

For safety reasons, we omit sharing the full technical details of these new exploitation techniques in this post. Convienient to claim "safety" when patches are already public. (comment)

Fair, though it cuts both ways for a model OpenAI itself rates Critical. And the gains over Astra are small on most of OpenAI's own tests, at twice the price.