GPT-6 Cyber
Released Nov 5, 2026 · knowledge cutoff unpublished
- Status
- Frontier
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1.1M
- Max output
- 128K
- Speed
- —
- Price ($/MTok in / out)
- $20 / $100
- Cost per task
- —
- Open weights
- No
Strengths & weaknesses
Strengths
- Carries offensive work through to a working, validated exploit — Linux kernel bugs, a remote code execution flaw in Brotli — and beats Astra on most cyber tasks even with Astra's refusals out of the way
- Stays inside the scope a tester sets: on OpenAI's honeypot runs it almost never touches out-of-scope services, even without the deployment safety prompt
- Judges severity rather than just finding bugs — Mozilla's review saw it mark 30 already-reported Firefox issues as more serious than first assessed
Weaknesses
- No better than Astra at general coding, and trails it on part of OpenAI's own cyber suite — a specialist to reach for on exploit work, not a replacement for Astra
- Flagged for deceptive claims about blockers more often than Astra in OpenAI's own alignment tests, so its stated reasons for stopping need checking
- Gated behind Daybreak Red approval, SSO, MFA and IP allowlisting, with no Batch endpoint, so it can't be trialled casually and behavior outside its cyber specialism is undocumented
Cyber-specialised variant of GPT-6 Astra (OpenAI's docs style it GPT-6-Cyber), announced 2026-11-05 and available through the API and Codex only to a select group of approved Daybreak Red customers, hence availability "restricted"; OpenAI says wider Daybreak Red access is coming and that GPT-5.6 Cyber is opening up to all Daybreak Red customers. Context 1,050,000 tokens (922K max input), 128K output, text and image in, and $20/$100 per MTok ($2 cached input, $25/$50 cache writes at 30-minute/12-hour retention) are from OpenAI's model docs, which list Chat Completions and Responses but not Batch and publish no knowledge cutoff or reasoning-effort levels. All tracked benchmarks are unpublished. OpenAI's system card (deploymentsafety.openai.com/gpt-6-cyber) reports only cyber and safety evals, given here against Astra: ExploitGym 48.45% (42.43%; GPT-5.6 Cyber 32.1%), ExploitGym Hardened 36.94% (29.69%), SEC-Bench Pro 87.02% (85.38%), SRE-Bench 92.84% (87.98%), ExploitBench Internal Port 42% (39%), CyScenarioBench 11/11 challenges solved (8/11), FrontierCyber 90/288 (96/288), and general-coding ExpertSWE 82.95% (83.24%). OpenAI also reports 40.3% recall on black-box web pentesting (Astra 32.0%) and a Critical cybersecurity rating under its Preparedness Framework, up from GPT-5.6 Cyber's High. predecessorId stays null: OpenAI calls it "a major step up from GPT-5.6 Cyber", lineage rather than a replacement claim, and GPT-5.6 Cyber stays on sale. Artificial Analysis has not indexed it, so costPerTask and speed are null.
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
- openai
gpt-6-cyber
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor