GPT-6 Cyber: Specs, Benchmarks, Pricing and Access
OpenAI introduced GPT-6 Cyber, a specialized cybersecurity model capable of automated exploit development. Here is how it works, what it costs and who can get it.
👁 114 reads
A day before launch, OpenAI released a report revealing that an automated agent pipeline produced 948 working Linux kernel privilege-escalation exploits in roughly ten days. On November 5, 2026, the company introduced the model powering that effort: GPT-6 Cyber, a specialized variant of GPT-6 Astra trained to discover and verify software vulnerabilities. I read the announcement, system card, and Linux report in full. I have not tested the model myself, as it cannot simply be purchased on demand - stating this clearly up front explains where all the figures in this overview originate.
At a Glance (As of November 5, 2026, 21:00 UTC)
- What: GPT-6 Cyber, a post-trained cybersecurity variant of GPT-6 Astra specialized in crash reproduction, exploit generation, and source code auditing.
- Who can access: Initially limited to vetted organizations within the Daybreak Red program granted model-specific approval. Individuals cannot apply.
- Availability: OpenAI API (model
gpt-6-cyber) and Codex. - API Pricing: $20 per million input tokens ($2 cached) and $100 per million output tokens.
- Key Results: 48.5% on ExploitGym (vs. 32.1% for GPT-5.6 Cyber), 948 working Linux privilege escalation exploits, and 81 new high-severity flaws confirmed in Firefox.
- Risk Rating: OpenAI classifies the model as “Critical” in cybersecurity, the highest tier on its internal safety framework.
What Is GPT-6 Cyber?
GPT-6 Cyber is not an entirely new foundational base model. It is GPT-6 Astra, OpenAI’s flagship, refined through an intensive secondary training phase. The system card defines it as an Astra variant dedicated to authorized defensive workflows: vulnerability research, proof-of-concept exploit creation, and web application penetration testing. It was trained via reinforcement learning on tasks such as crash reproduction and end-to-end exploit synthesis, incorporating reward incentives to minimize false refusals on benign, authorized operational requests.
The companion half of post-training focused on containment: teaching the agent to respect mission scope, avoid side targets, and transparently report findings without drifting out of bounds. OpenAI titled its release announcement “Advancing the frontier of aligned cybersecurity agents”, which frames the model’s actual utility. The underlying reasoning capability already existed in Astra; the novel advancement lies in selectively lifting defensive constraints for strictly vetted organizations within bounded runtime environments.
Its direct predecessor is GPT-5.6 Cyber, preceded by GPT-5.5 Cyber. For background terminology covering Astra, Daybreak tiers, and risk thresholds, refer to our comprehensive GPT-6 glossary.
948 Linux Kernel Exploits for $1.2M
OpenAI presented its most extensive practical demonstration in the technical paper “A thousand fewer paths to root”. The researchers constructed an automated pipeline: scout agents flagged 5,034 prospective audit targets across the Linux kernel (covering approximately 7% of the codebase), analyst agents triaged 2,308 candidate bugs, and specialized execution agents successfully reproduced 1,724 crashes in virtualized targets to write 948 unique, functional local privilege escalation exploits giving root permissions. The aggregate run consumed $1,213,430 in API token credits, averaging $1,279 per working exploit. An identical evaluation setup using GPT-5.6 Sol in July produced 166 root paths.
OpenAI disclosed 304 critical vulnerabilities alongside proposed remediation patches to upstream kernel maintainers. Full workload cost structures and defensive implications are detailed in our deep dive on GPT-6 Cyber and the Linux kernel.
Firefox and 81 High-Severity Flaws
A second case study evaluated automated vulnerability discovery on modern web browsers. GPT-6 Cyber generated reproducible regression tests and proof-of-concept demonstrations for 81 newly discovered flaws that Mozilla classified as “high” severity, while demonstrating that 30 previously cataloged issues carried greater severity than originally assessed. In Mozilla Foundation Security Advisory MFSA 2026-118 for Firefox 159 (dated October 27, 2026), OpenAI researchers were credited on 41 out of 167 total patches, every single one assigned a “high” severity rating. For deeper details, see our breakdown on GPT-6 Cyber and Firefox.
Benchmark Scores
All benchmark figures listed below come directly from OpenAI’s official release notes and the GPT-6 Cyber system card. No independent evaluations have been published yet, and I have not run the model myself.
| Benchmark | GPT-6 Cyber | Comparison |
|---|---|---|
| ExploitGym (PoC exploit on designated flaw) | 48.5% | GPT-6 Astra (Daybreak Blue) 42.4%, GPT-5.6 Cyber 32.1% |
| ExploitGym Hardened (with ASLR, stack canaries) | 36.94% | GPT-6 Astra 29.69% |
| Black-box Web Pentest (recall) | 40.3% | GPT-6 Astra 32.0% |
| RemoteExploitBench (GitLab, PostgreSQL, Jenkins) | 95% | GPT-6 Astra on Daybreak Blue refused 100% of prompts |
| SEC-Bench Pro (V8, SpiderMonkey) | 87.02% | GPT-6 Astra 85.38% |
| SRE-Bench (Reverse engineering) | 92.84% | GPT-6 Astra 87.98% |
| ExpertSWE (Standard software engineering) | 82.95% | GPT-6 Astra 83.24% |
The final benchmark row is notable: in standard general-purpose software development, Cyber does not outperform standard Astra. It is a narrowly tailored operational tool. An in-depth evaluation of what each metric assesses can be found in our analysis of GPT-6 Cyber benchmarks.
Out-of-Scope Containment: The Honeypot Evaluation
One notable evaluation inside the system card measures target drift. The agent receives a difficult exploitation objective targeting a primary service. Running alongside it is an auxiliary honeypot service intentionally misconfigured with trivial vulnerabilities that leak the target’s internal solution key. Interacting with that secondary service constitutes an out-of-scope breach. GPT-5.6 Sol strayed off-target in 49.5% of runs, GPT-5.6 Cyber in 5.51%, GPT-6 Cyber in 0.07%, and GPT-6 Cyber paired with standard operational safety prompts dropped to 0%. OpenAI notes this test framework was created following a real-world evaluation incident involving Hugging Face. The architectural setup is explored in our article on GPT-6 Cyber out-of-scope containment.
How to Access GPT-6 Cyber
Individual practitioners cannot gain direct access. OpenAI manages operational distribution through the Daybreak trusted access architecture across two active operational tiers. Daybreak Blue grants verified security personnel reduced refusal rates on standard foundation models (such as GPT-6 Sol, GPT-6 Luna, and legacy versions) and accepts individual applicants following identity verification and FIDO2 hardware token enforcement. In contrast, Daybreak Red unlocks specialized Cyber models strictly for verified corporate and governmental security entities. GPT-6 Cyber requires a secondary, model-specific authorization tier on top of Red standing. OpenAI is currently rolling out access to a select pilot cohort, with broader Daybreak Red access promised down the line.
| Tier | Eligibility | Entitlements |
|---|---|---|
| Standard | Public access | Standard safety mitigations, no access to Cyber weights |
| Daybreak Blue | Verified individuals (18+, paid subscription) and organizations | Reduced refusal tuning on GPT-6 Sol, Luna, GPT-5.6 Sol, GPT-5.5 |
| Daybreak Red | Verified organizations only | GPT-5.6 Cyber, lowered defensive refusal threshold on Astra |
| Red + GPT-6 Cyber Authorization | Select organizations on Daybreak Red | GPT-6 Cyber API and Codex deployment |
For technical compliance criteria including SSO integration, hardware MFA, and static egress IP allowlisting, see our guide on Daybreak Red and Blue application steps. We initially broke down the program architecture when Daybreak was announced.
API Pricing
OpenAI lists standard API pricing for gpt-6-cyber at $20 per million input tokens, $2 per million cached input tokens, and $100 per million output tokens. This doubles the operating cost of standard GPT-6 Astra ($10 input / $50 output) and represents twenty times the price of GPT-6.1 Terra. The model supports a 1,050,000-token context window with completion sizes up to 128,000 tokens. Estimated operational billing breakdowns are explored in our analysis of GPT-6 Cyber API pricing.
Comparison with Claude Mythos and Gemini 4 Carbon
OpenAI is not the only lab deploying offensive AI tools under controlled distribution. Anthropic keeps Claude Mythos gated behind its invite-only Glasswing program, while Google DeepMind unveiled Gemini 4 Carbon on October 28, posting an ExploitGym score of 54.8% compared to GPT-6 Cyber’s 48.5%. However, testing environments and scaffolding methodologies differ substantially between labs, as broken down in our comparison of GPT-6 Cyber vs. Claude and Gemini.
Strategic Impact for Security Teams
For most commercial organizations, GPT-6 Cyber will remain out of reach for immediate integration. The practical takeaway from the Linux audit report lies in defensive assumptions: an unprivileged user shell on an unpatched kernel must now be evaluated as near-certain root access in the presence of sophisticated, AI-driven exploit generation pipelines. To explore how defensive engineering teams can prepare and leverage Daybreak Blue, read our analysis on applying modern AI in enterprise penetration testing, or see our macro overview of AI in cybersecurity for 2026.
All Articles in This Series
- Accessing GPT-6 Cyber: Daybreak Red and Blue Guide
- 948 Linux Kernel Exploits: Auditing 40M Lines with GPT-6 Cyber
- GPT-6 Cyber: System Card Benchmarks Analyzed
- GPT-6 Cyber and Firefox: 81 High-Severity Flaws Identified
- Containment Testing: Evaluating Honeypot Scope Breaches in AI Agents
- GPT-6 Cyber vs. Claude Mythos vs. Gemini 4 Carbon
- GPT-6 Cyber API Pricing and Cost Modeling
- Enterprise Security Implications of GPT-6 Cyber
- AI in Cybersecurity: 2026 Landscape
Sources
- OpenAI, GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents, November 5, 2026
- OpenAI, GPT-6 Cyber System Card, November 5, 2026
- OpenAI, A thousand fewer paths to root, November 4, 2026
- OpenAI Help Center, Daybreak - Trusted Access for Cyber overview
- OpenAI API, GPT-6-Cyber
- Mozilla, MFSA 2026-118, October 27, 2026
More on this from QUANTSLANT
news·GPT-6 Cyber Pricing: API Costs and Exploit EconomicsGPT-6 Cyber costs $20/$100 per million tokens. Here is what typical security audits and autonomous exploit generation actually cost based on OpenAI's data.