GPT-5.6-Cyber
Released Aug 10, 2026 · knowledge cutoff 2026-02
- Status
- Frontier
- Location
- United States
- Modality
- Multimodal
- Context window
- 400K
- Max output
- 128K
- Speed
- —
- Price ($/MTok in / out)
- $12.50 / $75
- Cost per task
- —
- Open weights
- No
Benchmarks0 of 8 reported
MMLU-Pro
GPQA Diamond
SWE-bench Verified
Terminal-Bench 2.1
AIME (latest)
Humanity's Last Exam
LMArena Elo
ARC-AGI-2
Strengths & weaknesses
Strengths
- Carries out the dual-use half of security work — exploit reproduction, authentication bypass, privilege escalation — that the general-purpose Sol build refuses outright
- Holds a vulnerability hunt together long enough to surface undocumented bugs in heavily audited code, including a chained V8 sandbox escape in Chrome
- Takes screenshots and diagrams alongside source, so binary and tooling output doesn't have to be transcribed to text first
Weaknesses
- Writes shorter, thinner vulnerability reports than the general-purpose models and is weaker across broad security assessment — a specialist to reach for, not a replacement for Sol on review work
- Every request sits behind identity verification, monitoring, approved-use restrictions and legal attestations, so it can't be trialled or adopted casually
- No published general-capability scores and no independent leaderboard placement — behavior outside its cyber specialism is undocumented
Purpose-trained cybersecurity variant built on GPT-5.6 Sol, reachable only through Daybreak Red, the vetted upper tier of OpenAI's Daybreak defender programme (Daybreak Blue covers general-purpose models), hence availability "restricted". Context window (400K, 272K max input), max output, knowledge cutoff (2026-02-16) and pricing ($12.50/$75 per MTok, cached input $1.25) are from OpenAI's own model docs; the model is served on the Responses API only. All eight tracked benchmarks are unpublished — OpenAI reported only cyber-specific evals: a 95.0% Advanced Cybersecurity Completion Rate against 57.3% for GPT-5.5-Cyber and 1.5% for GPT-5.6 Sol, plus an ExploitGym lead over both, and a Preparedness Framework cybersecurity rating of High but below Critical. Artificial Analysis has not indexed it, so costPerTask and speed are null, and its vetted-access gating makes future indexing unlikely. Announced 2026-08-10; OpenAI's API changelog listed the model and the daybreak-red-latest alias on 2026-08-07. GPT-5.5-Cyber (June 2026) is not tracked here and OpenAI states no supersession, so predecessorId stays null.
For developers
API model strings
- openai
gpt-5.6-cyber
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor