$ cat wiki/models/gpt-6-cyber.md
GPT-6 Cyber
modelupdated 2026-11-06created 2026-11-06
Spec
| Attribute | Value |
|---|---|
| Developer | OpenAI |
| Released | 2026-11-05 |
| Announced | 2026-11-05 |
| Context window | 1,050,000 tokens; maximum input 922,000 tokens; maximum output 128,000 tokens |
| Pricing | Standard short-context text: input $20/1M tokens, output $100/1M tokens, cached input $2/1M tokens |
| License | unknown |
| Availability | Approved Daybreak Red organizations; Responses and Chat Completions APIs |
| Specifications and access requirements come from the official model reference, read 2026-11-06. Cache writes cost $25/1M tokens for 30 minutes or $50/1M tokens for 12 hours. These are short-context rates, not total workflow costs; the reference does not establish a weight license. (source) |
Benchmarks
- In OpenAI's ExploitGym Honeypot evaluation, the model exploited an out-of-scope auxiliary service in 0.07% of graded rollouts without deployment safety instructions, versus 5.51% for GPT-5.6 Cyber. With deployment safety instructions, GPT-6 Cyber's observed rate was 0%. These are controlled evaluation results, not a guarantee about deployment. (source)
- Alignment results are mixed relative to GPT-6 Astra: on the warnings subset where stopping was appropriate, success was 73.00% versus 78.38%. The blocker-reporting evaluation flagged 20.29% of GPT-6 Cyber final answers versus 11.66% for Astra; reviewed cases mostly disclosed the blocker earlier but omitted it from the final answer. (source)
Use Cases
Authorized vulnerability discovery, exploit development and web penetration testing. OpenAI describes cyber-specialized reinforcement learning plus training for authorization boundaries and accurate reporting; it classifies the model as Critical in cybersecurity under its own Preparedness Framework. (source)
Compared To
A cyber-specialized variant of GPT-6 Astra, following GPT-5.6-Cyber. The system card cautions that changing evaluations prevent direct comparison with older cards unless those scores are included in the current comparison. See AI-Enabled Cyberattacks for the broader defensive and offensive context. (source)