GPT-6 Cyber wins authorized exploit work outright. OpenAI launched it on November 5, 2026 as a cybersecurity variant of GPT-6 Astra, and on OpenAI's own web pentesting eval it recalled 40.3% of vulnerabilities against 32.0% for Astra. On RemoteExploitBench, Astra with Daybreak Blue refused every task while GPT-6 Cyber completed 95% (both figures from OpenAI's launch post). For defensive work Astra remains the model you can actually reach, because GPT-6 Cyber is gated behind vetting.
TL;DR
- GPT-6 Cyber scored 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber, per OpenAI.
- It beats GPT-6 Astra on pentest recall (40.3% vs 32.0%) and on completing known-exploit tasks, per OpenAI.
- Access: approved Daybreak Red customers only; identity verification, SSO, MFA and IP allowlisting required.
- List price: $20 per 1M input and $100 per 1M output tokens, per the OpenAI API docs.
- OpenAI rates it Critical under its Preparedness Framework, up from High.
- Choose GPT-6 Cyber for sanctioned offensive research if you qualify; GPT-6 Astra for defensive work.
How does GPT-6 Cyber compare with GPT-6 Astra on exploit work?
GPT-6 Cyber is post-trained for vulnerability discovery, exploit development and web pentesting. Table figures come from OpenAI's launch post and the system card.
| Capability test | GPT-6 Cyber | GPT-6 Astra | GPT-5.6 Cyber |
|---|---|---|---|
| ExploitGym (exploit a given flaw) | 48.5% | not reported | 32.1% |
| Web pentest vulnerability recall | 40.3% | 32.0% | not reported |
| RemoteExploitBench completed | 95% | refused 100% | not reported |
| Honeypot stayed in scope (no deployment prompts) | 99.93% | not reported | 94.49% |
| Preparedness rating (cyber) | Critical | High | High |
The ExploitGym jump is 16.4 percentage points (48.5% minus 32.1%), a 51% relative improvement over GPT-5.6 Cyber, per OpenAI. RemoteExploitBench uses known vulnerabilities in GitLab, PostgreSQL and Jenkins in a controlled environment. Astra's 100% refusal is by design: a general model's safeguards, not a failure.
Is GPT-6 Cyber actually finding real vulnerabilities?
OpenAI reports these real-world results, all from the launch post:
- Linux privilege escalation: 948 working exploits from 1,724 reproduced vulnerabilities, drawn from 2,308 candidates. GPT-5.6 Sol produced 166 on the same harness, so GPT-6 Cyber's output is more than five times larger (948 against 166). OpenAI says 304 critical findings were disclosed to maintainers.
- Firefox: 81 novel high-severity vulnerabilities with proposed patches. Mozilla has published fixes in advisory mfsa2026-118.
- Brotli: a critical remote code execution flaw, CVE-2026-111497, CVSS 9.2, described at patchtheplanet.org.
The Mozilla advisory is the strongest evidence: a third party shipped fixes. BleepingComputer and CyberScoop repeat the figures but do not reproduce them.
How much does GPT-6 Cyber cost per token?
The OpenAI API docs list the prices below. INR figures use about 94.05 rupees per dollar, the Xe rate on November 8, 2026, and will drift.
| Per 1M tokens | USD | INR (approx.) |
|---|---|---|
| Input | $20 | Rs 1,881 |
| Cached input | $2 | Rs 188 |
| Output | $100 | Rs 9,405 |
Cache writes cost $25 per 1M tokens for a 30-minute window or $50 for 12 hours. The context window is 1,050,000 tokens, with a maximum input of 922,000 and maximum output of 128,000, per the OpenAI API docs. It supports reasoning tokens and accepts text and image input.
This is the highest price OpenAI publishes; VentureBeat reports it sits above Anthropic's and Google's cyber models. Claude Fable 5 lists at $10 input and $50 output per 1M tokens, half the rate. Our GPT-6 Astra vs Claude Fable 5.1 verdict covers the general-purpose side.
How do you get access to GPT-6 Cyber?
You cannot sign up and call it. The model ID gpt-6-cyber initially runs through the API and Codex for approved Daybreak Red customers, per the launch post. The steps:
- Apply to Daybreak Red at openai.com/daybreak.
- Pass identity verification.
- Configure SSO and MFA for your organisation.
- Provide IP allowlisting for the addresses that will call the API.
- Wait for approval. OpenAI says it will progressively expand to all Daybreak Red customers.
The requirements come from the system card. OpenAI told VentureBeat it had enrolled upwards of 2,000 entities in Daybreak Red last month. GPT-5.6 Cyber is also opening to all Daybreak Red customers, the practical fallback while you wait. Defenders wanting Astra with fewer restrictions can join the Daybreak Blue alpha via the form in the launch post.
One data-handling change matters for individuals: from November 29, 2026, OpenAI plans to use content from Daybreak Blue sessions on personal accounts to improve its models even when the "improve the model for everyone" setting is off, per The Hacker News. Consider that before routing client material through a personal account.
Which model should you choose?
Choose GPT-6 Cyber if you run sanctioned penetration tests, exploit validation or vulnerability research and can clear Daybreak Red vetting. Its measured lead on exploit tasks and its 99.93% in-scope behaviour on the honeypot eval (up from 94.49% for GPT-5.6 Cyber, per OpenAI) are the reasons.
Choose GPT-6 Astra for defensive work, triage, code review or anything outside exploit development: it is far easier to obtain and its refusals are expected behaviour for a general model. If your work is ordinary software development, our comparison of the best LLMs for coding is more relevant than any cyber-specific model.
The main caveat is risk: OpenAI itself rates the model Critical. Dragos chief AI officer Rob T. Lee told SecurityWeek that the skill floor is dropping for pre-attack reconnaissance and vulnerability research. OpenAI also announced a $1 billion Daybreak for Frontline Defenders grant programme and a Daybreak Red for Critical Infrastructure group, per the same report. Related: OpenAI's quadrupled critical bug bounty payouts and GroundingPi XPeng 4B against GPT-6 Astra.
FAQ
Q: Can Indian developers get access to GPT-6 Cyber?
A: OpenAI has not published a country restriction in the sources we checked, but access runs through Daybreak Red: identity verification, SSO, MFA and IP allowlisting. Indian security firms can apply at openai.com/daybreak; approval is at OpenAI's discretion.
Q: Is GPT-6 Cyber available in ChatGPT?
A: The launch post lists the API and Codex as the initial access routes for approved Daybreak Red customers. It does not describe a general ChatGPT release.
Q: How much does a security review cost with GPT-6 Cyber?
A: It depends on token volume. As an illustration, 2M uncached input tokens plus 300,000 output tokens costs $70 (about Rs 6,584) at the published rates; cached input at $2 per 1M lowers repeated-context costs. The token counts are our assumption, not a measured workload.
Q: Does GPT-6 Cyber beat GPT-5.6 Cyber?
A: On ExploitGym, yes: 48.5% against 32.1%, per OpenAI. That is vendor-reported, and GPT-5.6 Cyber is the model more customers can reach first.
Q: Why did GPT-6 Astra refuse RemoteExploitBench?
A: Astra with Daybreak Blue refused all tasks targeting known vulnerabilities in GitLab, PostgreSQL and Jenkins; its safeguards decline exploit development, while GPT-6 Cyber completed 95%, per OpenAI.
Last verified: November 8, 2026.
This article was drafted with AI assistance and reviewed against the cited primary sources. Benchmark figures are reported by the vendor and have not been independently reproduced.
Top comments (0)