GPT-5.6-Cyber
Purpose-trained model for advanced cybersecurity research and vulnerability testing
Context window
400K
Input / 1M tokens
$12.50
Output / 1M tokens
$75.00
Provider
OpenAI
Cached input $1.25 per 1M; cache writes $15.625 (30-minute) or $31.25 (12-hour), per the Cyber models section of OpenAI's pricing page. The context window is 400K tokens (272K max input, 128K max output). Only the Responses API is supported; Batch is not. GPT-6 Cyber, added November 5, 2026, costs $20.00/$100.00 with a 1.05M context window. · Data verified 2026-11-08
GPT-5.6-Cyber is a specialized variant of GPT-5.6 Sol trained with reduced safeguards to excel at cybersecurity tasks including vulnerability research, exploit chain development, authentication bypass, and privilege escalation. Built on the GPT-5.6 foundation, it completed 95% of advanced cybersecurity requests in testing, significantly outperforming GPT-5.6 Sol (1.5%) and earlier GPT-5.5-Cyber (57.3%).
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Available to Daybreak Red customers through the OpenAI Responses API only (model id gpt-5.6-cyber), with Daybreak Red enabled for the project that sends requests. Applicants must verify lawful, defensive, authorized work and undergo vetting through Daybreak Access. Since November 4, 2026 approved customers can also use it in AWS GovCloud (US-East) through Amazon Bedrock, which needs both Daybreak Red approval and AWS provisioning.
Strengths
- ✓95% completion rate on advanced cybersecurity tasks
- ✓Enhanced capability for exploit chain development and validation
- ✓Ability to identify zero-day vulnerabilities in real-world software (Chrome V8, operating systems)
- ✓Superior performance on ExploitGym benchmarks vs. prior versions
Best for developers who...
When to choose it (and when not to)
Reach for GPT-5.6-Cyber when...
- →Organizations need advanced cybersecurity capabilities with reduced refusals
- →Conducting authorized vulnerability research on owned or permitted systems
- →Testing complex multi-step attack scenarios
Look elsewhere if...
- ✕General-purpose tasks (use GPT-6.1 Sol or GPT-6 Astra)
- ✕Exploratory security work without authorization (use Daybreak Blue tier)
- ✕If not accepted into OpenAI's restricted Daybreak Red program
- ✕You have access to GPT-6 Cyber (November 2026), OpenAI's newer cyber model
- ✕You need more than 272K input tokens (GPT-6 Cyber accepts up to 922K)
How to use it
- ›Clearly specify the scope and authorization for security testing
- ›Include details of target systems and objectives for best results
- ›Use structured prompts for exploit chain development
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI() # key with Daybreak Red access
response = client.responses.create(
model="gpt-5.6-cyber",
input="Outline an assessment plan for this authorized test scope: ...",
)
print(response.output_text)GPT-5.6-Cyber works only through the Responses API and requires Daybreak Red access.
API model id: gpt-5.6-cyber
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Advanced Cybersecurity Task Completion | 95% | Completion rate on complex requests including exploit chains, authentication bypass, privilege escalation |
| OpenAI Preparedness Framework - Cybersecurity Capability | High | Assessed as High for cybersecurity capability but below Critical threshold |
Compare GPT-5.6-Cyber
Compare GPT-5.6-Cyber with any other model
Build a comparison →All model comparisons →