OpenAI released GPT-6 Cyber on November 5, 2026, introducing a specialized model built specifically for authorized cybersecurity work. According to OpenAI’s system card, GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra trained to discover and validate exploitable vulnerabilities within boundaries defined by security researchers.
Unlike general-purpose releases in the fall 2026 AI model wave, GPT-6 Cyber is not a consumer chatbot upgrade. It is gated behind OpenAI’s Daybreak Red access tier and is currently rolling out via the API (gpt-6-cyber) and Codex to approved security organizations.
Here is what OpenAI announced, how GPT-6 Cyber compares to GPT-6 Astra and GPT-5.6 Cyber, what it costs, and what your organization needs in order to access it.
What is GPT-6 Cyber?
GPT-6 Cyber is a post-trained variant of GPT-6 Astra engineered for offensive-grade, defender-authorized security tasks such as vulnerability discovery, exploit development, and web penetration testing.
General frontier models often refuse legitimate penetration testing prompts because safety filters cannot easily tell an authorized researcher apart from an attacker. According to OpenAI, GPT-6 Cyber received additional post-training to improve its raw cybersecurity capability while reducing unnecessary refusals on authorized security requests.
At the same time, OpenAI trained the model to stay strictly inside researcher-defined scopes and report findings accurately. In its announcement, OpenAI said the model surpasses GPT-5.6 Cyber at finding and fixing vulnerabilities while remaining more reliably inside designated target boundaries.
Under OpenAI’s Preparedness Framework, both GPT-6 Cyber and GPT-6 Astra carry a Critical designation in Cybersecurity, up from the High rating assigned to GPT-5.6 Cyber. OpenAI’s system card also lists GPT-6 Cyber as High in the Biological and Chemical category and below High in AI Self-Improvement, matching Astra’s designations.
How does GPT-6 Cyber differ from GPT-6 Astra?
GPT-6 Cyber differs from GPT-6 Astra primarily in its refusal tuning for offensive validation tasks, its stronger benchmark performance on security evaluations, and its restricted Daybreak Red access path.
While mainline models in the GPT-6 Astra, Sol, and Luna lineup already handle code review, vulnerability triage, and patch development, they are built to refuse deeper exploit generation. OpenAI pairs GPT-6 Astra with Daybreak Blue—an access tier for verified defenders doing vulnerability and patch validation with fewer false-positive refusals. GPT-6 Cyber is paired with Daybreak Red, which supports deeper exploit development and authorized security testing.
OpenAI also states that GPT-6 Cyber outperforms GPT-6 Astra on most cybersecurity tasks even when Astra does not refuse the prompt.
| Model | Primary Role | Preparedness Cyber Tier | Access Path | Notable OpenAI Benchmark Scores |
|---|---|---|---|---|
| GPT-6 Cyber | Deep vulnerability discovery, exploit validation, and authorized web pentesting | Critical | Daybreak Red (approved organizations via API and Codex) | 48.5% on ExploitGym; 40.3% recall on black-box web pentesting; 95% completion on RemoteExploitBench |
| GPT-6 Astra | Flagship general reasoning, code review, vulnerability triage, and patch development | Critical | Standard API / ChatGPT tiers; Daybreak Blue alpha for verified defenders | 32.0% recall on black-box web pentesting; 100% refusal on RemoteExploitBench under Daybreak Blue |
| GPT-5.6 Cyber | Previous-generation specialized security model | High | Available to all approved Daybreak Red customers | 32.1% on ExploitGym; 94.49% out-of-scope avoidance on ExploitGym Honeypot (without deployment prompt) |
According to OpenAI’s published evaluations, the gap between the models shows up clearly in controlled testing:
- ExploitGym: OpenAI reports that GPT-6 Cyber reached 48.5%, compared to 32.1% for GPT-5.6 Cyber—an improvement of roughly 16 percentage points, or about 51% relative improvement according to the announcement.
- Black-box web pentesting: On source-free web penetration evaluations, OpenAI says GPT-6 Cyber achieved 40.3% recall versus 32.0% for GPT-6 Astra.
- RemoteExploitBench: In controlled testing against known vulnerabilities in software such as GitLab, PostgreSQL, and Jenkins, OpenAI says GPT-6 Astra under Daybreak Blue refused 100% of requests, whereas GPT-6 Cyber completed 95% (with the single failure resulting from an unsuccessful remote exploit attempt rather than a refusal).
- ExploitGym Honeypot (scope alignment): Without a deployment safety prompt, GPT-6 Cyber avoided attacking out-of-scope services 99.93% of the time, up from 94.49% for GPT-5.6 Cyber. With the deployment safety prompt enabled, OpenAI reports 100% compliance.
What can approved teams use GPT-6 Cyber for?
Approved security teams can use GPT-6 Cyber to audit large codebases, validate whether candidate bugs are genuinely exploitable, run scoped web application tests, and draft candidate patches for maintainers.
In high-volume security engineering, teams often burn hours chasing static analysis alerts that cannot actually be triggered in production. By validating whether a flaw is exploitable inside an isolated test harness and generating a candidate fix, a specialized agent helps defenders prioritize real risk.
OpenAI shared two major real-world hardening projects in its November 5 announcement:
- Linux kernel hardening: According to OpenAI, researchers used the model in controlled harnesses to reproduce 1,724 vulnerabilities and develop 948 working privilege-escalation exploits—compared to 166 under the same harness using GPT-5.6 Sol. OpenAI says it disclosed 304 critical findings alongside candidate patches to maintainers.
- Firefox security audits: OpenAI reported discovering 81 previously unknown vulnerabilities that Mozilla rated as high severity, while showing that another 30 already-known issues were more severe than previously understood. According to OpenAI, Mozilla has already begun merging fixes.
OpenAI’s launch announcement also included commentary and early testing references from organizations across security and developer tooling, including Cisco, Calif, CodeRabbit, Palo Alto Networks, SentinelOne, Perplexity, and Cognition. Because defenders must also guard their own AI pipelines against adversarial inputs, pairing code audits with defenses against prompt injection remains essential for teams deploying agentic workflows.
Who can access GPT-6 Cyber and Daybreak Red?
Access to GPT-6 Cyber is restricted to vetted organizations approved for OpenAI’s Daybreak Red program, starting with an initial cohort on the API and in Codex before expanding to all Daybreak Red customers.
Individual ChatGPT Plus or Pro subscribers cannot simply toggle GPT-6 Cyber on in their consumer accounts. Instead, OpenAI splits its security access programs into two tracks:
- Daybreak Blue (Alpha): Aimed at verified defenders and expanding to more organizations and individual security researchers, using models like GPT-6 Astra for defensive vulnerability and patch validation.
- Daybreak Red: Designed for organizations conducting deeper exploit research and authorized security testing. While GPT-6 Cyber rolls out gradually to this group, OpenAI says all Daybreak Red members immediately receive access to GPT-5.6 Cyber. OpenAI also introduced a Daybreak Red for Critical Infrastructure cohort for approved critical infrastructure defenders.
According to OpenAI’s system card and model documentation, getting approved for Daybreak Red requires strict organizational and technical controls:
- Identity verification and account approval: Organizations must apply for Daybreak and have their eligibility confirmed through their OpenAI account team.
- Mandatory security posture: Participating organizations must enforce Single Sign-On (SSO), Multi-Factor Authentication (MFA), and API IP allowlisting.
- Private Safety Processing (PSP) for Zero Data Retention: Customers using Zero Data Retention (ZDR) are generally required to enable Private Safety Processing for automated safety classification. Under PSP, encrypted records remain in customer-controlled storage for roughly 30 days, and OpenAI personnel do not view prompts or responses.
How much does the GPT-6 Cyber API cost?
The gpt-6-cyber API is priced at $20 per million input tokens and $100 per million output tokens for standard short-context requests, making it a premium tier well above standard coding and chat models.
According to OpenAI’s official API model page (verified November 6, 2026), the full token pricing and technical specifications are:
- Model ID:
gpt-6-cyber - Input tokens: $20 per 1M tokens
- Cached input tokens: $2 per 1M tokens
- Cache writes: $25 per 1M tokens
- Output tokens: $100 per 1M tokens
- Context window: 1,050,000 tokens (up to 922,000 maximum input tokens and 128,000 maximum output tokens, with reasoning token support)
- Modalities: Text and image inputs; text outputs
On the integration side, gpt-6-cyber supports both the Responses API and Chat Completions API. It includes support for streaming, structured outputs, function calling, web search, and prompt caching, plus Responses API tools including code_interpreter, hosted_shell, apply_patch, computer_use, file_search, tool_search, image_generation, skills, and mcp. The Batch API is not supported.
To put those numbers in perspective, our AI API pricing comparison shows that mainstream flagship models cost a fraction of GPT-6 Cyber’s per-token rate. That gap reflects both the heavy reasoning compute involved in multi-step security agents and the compliance overhead of Daybreak Red.
What should security and product teams do next?
Most software teams should keep using GPT-6 Astra or GPT-6 Sol for everyday code review and reserve GPT-6 Cyber for dedicated offensive security and vulnerability research groups.
If your goal is general application development, automated PR reviews, or routine bug triage, you do not need Daybreak Red—and paying $100 per million output tokens will quickly drain your budget. Walk through our guide on how to choose an AI model to match your daily engineering tasks to a cost-effective tier, or use specialized creative platforms like AI Studio Video when your workflow is focused on media generation rather than security testing.
If you run a dedicated product security, red team, or critical infrastructure defense group:
- Decide between Daybreak Blue and Daybreak Red: If your primary hurdle is GPT-6 Astra refusing benign patch validation or defensive triage prompts, applying for the expanding Daybreak Blue alpha is the simpler path.
- Prepare your infrastructure for Daybreak Red: If your team conducts authorized exploit validation and needs
gpt-6-cyber, verify that your workspace meets OpenAI’s SSO, MFA, IP allowlisting, and Private Safety Processing requirements before submitting your application. - Use prompt caching aggressively: At $2 per million cached input tokens versus $20 for uncached inputs, caching static repository context and system instructions is essential for keeping automated audit costs predictable.
FAQ
What is OpenAI GPT-6 Cyber?
GPT-6 Cyber is a cybersecurity-specialized variant of GPT-6 Astra released by OpenAI on November 5, 2026. It is post-trained to help authorized security researchers discover vulnerabilities, validate whether bugs are exploitable, and run scoped penetration tests while staying strictly within researcher-defined boundaries and minimizing false-positive refusals.
Can ChatGPT Plus or Pro users access GPT-6 Cyber?
No. GPT-6 Cyber is not available to general consumer or standard business ChatGPT accounts. Access is limited to vetted organizations approved for OpenAI’s Daybreak Red program, where it is rolling out through the OpenAI API (gpt-6-cyber) and Codex under strict identity, SSO, MFA, and IP allowlisting requirements.
How much does GPT-6 Cyber cost on the OpenAI API?
On the OpenAI API, gpt-6-cyber costs $20 per million input tokens, $2 per million cached input tokens, $25 per million tokens for cache writes, and $100 per million output tokens. It supports a 1,050,000-token context window (up to 922,000 input and 128,000 output tokens) and does not support Batch API discounts.
What is the difference between Daybreak Blue and Daybreak Red?
Daybreak Blue is OpenAI’s program for verified defenders using mainline models like GPT-6 Astra to validate vulnerabilities and patches with fewer unnecessary refusals. Daybreak Red is a higher-trust tier for organizations conducting deeper exploit development and authorized offensive security testing, granting access to GPT-6 Cyber and GPT-5.6 Cyber.
Is GPT-6.2 Astra released alongside GPT-6 Cyber?
No. As of November 6, 2026, OpenAI has not officially announced or released a model named GPT-6.2 Astra in its public changelog or API documentation. GPT-6 Astra remains OpenAI’s current flagship general-purpose model, while GPT-6 Cyber is its newly launched cybersecurity-specialized variant.
Sources
- OpenAI, GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents, November 5, 2026
- OpenAI Developer Documentation, GPT-6 Cyber API Model Page (
gpt-6-cyber), accessed November 6, 2026 - OpenAI Deployment Safety, GPT-6 Cyber System Card (PDF), November 5, 2026
- OpenAI, Product Releases News Index, November 5, 2026
- The Verge, OpenAI launches GPT-6 Cyber, November 5, 2026
- VentureBeat, OpenAI launches GPT-6 Cyber with stronger exploit research and higher API price, November 5, 2026
