GPT-6 Cyber
OpenAI's cybersecurity variant of GPT-6 Astra, limited to approved Daybreak Red organizations
Context window
1.05M
Input / 1M tokens
$20.00
Output / 1M tokens
$100.00
Provider
OpenAI
Cached input $2.00 per 1M; cache writes $25.00 (30-minute) or $50.00 (12-hour). Listed under Cyber models on OpenAI's pricing page, next to GPT-5.6-Cyber at $12.50/$75.00. Max input is 922,000 tokens and max output 128,000. The Batch API is not supported. · Data verified 2026-11-11
GPT-6 Cyber (November 5, 2026) is a cybersecurity variant of GPT-6 Astra for finding exploitable vulnerabilities and developing fixes, which OpenAI says is trained to stay within authorized scope. Access requires approval for OpenAI's Daybreak Red program, with identity verification and use restrictions. In OpenAI's system card it beats GPT-6 Astra on ExploitGym (48.45% vs 42.43%) and SEC-Bench Pro (87.02% vs 85.38%), and it is rated Critical for cybersecurity under OpenAI's Preparedness Framework, the same as Astra. API pricing is $20/$100 per million tokens with a 1.05M-token context window.
How to access it
Restricted: available through the OpenAI API (Chat Completions and Responses, model ID gpt-6-cyber) only to organizations with Daybreak Red access and approval. Organizations need single sign-on, multi-factor authentication and API IP allowlisting; invited individuals go through identity and liveness checks and must use a FIDO2 security key.
Strengths
- ✓48.45% on ExploitGym vs 42.43% for GPT-6 Astra (OpenAI system card)
- ✓Exploited an out-of-scope honeypot service in 0.07% of runs without deployment safety instructions, vs 5.51% for GPT-5.6 Cyber, and 0% with them (OpenAI system card)
- ✓1.05M-token context window with up to 128K output tokens
- ✓Responses API tools include function calling, web search, computer use, hosted shell and MCP
Best for developers who...
When to choose it (and when not to)
Reach for GPT-6 Cyber when...
- →Your organization is approved for Daybreak Red and wants OpenAI's newest cyber model
- →You need to reproduce vulnerabilities and verify fixes in an authorized test environment
Look elsewhere if...
- ✕General-purpose tasks (use GPT-6 Astra or GPT-6.1 Sol)
- ✕Your organization is not approved for Daybreak Red
- ✕Cost-sensitive work ($20/$100 per 1M tokens)
- ✕You need the Batch API (not supported)
How to use it
- ›State the authorized scope and test environment explicitly; OpenAI trained the model to stay within authorized scope
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI() # organization needs Daybreak Red approval
response = client.responses.create(
model="gpt-6-cyber",
input="Reproduce the reported heap overflow in our authorized test build and propose a patch: ...",
)
print(response.output_text)Responses API. Complete the organization and project steps in OpenAI's Daybreak guide before the first request.
API model id: gpt-6-cyber
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| ExploitGym | 48.45% | Intended-exploit success, OpenAI system card; GPT-6 Astra 42.43% |
| ExploitGym Hardened | 36.94% | OpenAI system card; GPT-6 Astra 29.69% |
| SEC-Bench Pro | 87.02% | pass@1, OpenAI system card; GPT-6 Astra 85.38% |
| SRE-Bench | 92.84% | pass@1, OpenAI system card; GPT-6 Astra 87.98% |
| ExploitGym Honeypot (out-of-scope exploitation) | 0.07% | Lower is better; 0% with deployment safety instructions; GPT-5.6 Cyber 5.51% |
Source: OpenAI - GPT-6 Cyber system card
Compare GPT-6 Cyber
Compare GPT-6 Cyber with any other model
Build a comparison →All model comparisons →