GPT-6 Cyber, announced by OpenAI on November 5, 2026, is the better model than GPT-5.6 Cyber for approved security teams: OpenAI reports it solves 48.5% of ExploitGym tasks against 32.1% for its predecessor, according to the OpenAI launch post. It costs more, at $20 per 1M input tokens against $12.50, and it is available only to a select group of approved Daybreak Red customers. If you are not in that group, stay on GPT-6 Astra, which remains the general-purpose choice.
TL;DR
- GPT-6 Cyber is a cybersecurity model built to find and validate exploitable vulnerabilities while staying inside the scope a researcher sets (OpenAI).
- On ExploitGym it scores 48.5% against 32.1% for GPT-5.6 Cyber. All benchmark figures are self-reported by OpenAI.
- API pricing is $20.00 input and $100.00 output per 1M tokens (OpenAI pricing).
- Access requires Daybreak Red approval. It is available in the API and Codex, and OpenAI says access will expand progressively.
- OpenAI rates the model Critical for cybersecurity under its Preparedness Framework (system card).
- Verdict: approved defenders should move to it; everyone else should use GPT-6 Astra.
What is GPT-6 Cyber and who is it for?
GPT-6 Cyber is a specialised model for defensive security work: finding vulnerabilities, writing proof-of-concept exploits to confirm they are real, and developing patches. OpenAI says it was trained to keep to the scope a researcher defines.
It is distributed through the API and Codex to approved Daybreak Red customers, per the launch post. OpenAI is also opening GPT-5.6 Cyber to all Daybreak Red customers, so there is now a wider tier and a narrower, more capable one.
How does GPT-6 Cyber perform on benchmarks?
OpenAI publishes the results below in its launch post. They are vendor-reported and have not been reproduced independently by us.
| Benchmark | GPT-6 Cyber | GPT-5.6 Cyber | GPT-6 Astra |
|---|---|---|---|
| ExploitGym (exploit a specified vulnerability on a remote target) | 48.5% | 32.1% | not reported |
| Web pentesting vulnerability-discovery recall | 40.3% | not reported | 32.0% |
| RemoteExploitBench (known GitLab, PostgreSQL and Jenkins flaws) | 95% of tasks completed | not reported | refused 100% of tasks (with Daybreak Blue) |
| ExploitGym Honeypot, stayed in scope without deployment prompts | 99.93% | 94.49% | not reported |
The ExploitGym gap is about 16 percentage points, a 51% relative improvement (OpenAI; corroborated by The Verge). The RemoteExploitBench row needs context: Astra's score reflects refusals, not inability. OpenAI also says GPT-6 Cyber outperforms Astra on most cyber tasks even when Astra does not refuse.
With the deployment safety prompt, OpenAI reports 100% in-scope behaviour in its launch post, so the prompt still does real work. Do not assume the model alone is a sufficient guardrail.
What has it found in real software?
OpenAI says its researchers used GPT-6 Cyber to build working exploits for 948 of 1,724 reproduced Linux kernel vulnerabilities, with GPT-5.6 Sol managing 166 on the same harness, per the launch post. The model also identified 81 previously unknown high-severity Firefox vulnerabilities and showed that 30 known issues were more severe than reported. Mozilla has released fixes for many of them. Source: OpenAI; the Firefox count is also reported by MarketScale.
For an independent view, CodeRabbit ran its own test through Daybreak Defense Network access. On 33 benchmark cases, its model ensemble with GPT-6 Cyber caught 22 selected issues against 19 without it, a 9.1% gain (CodeRabbit). It also raised 361 high- and medium-severity findings against 218. CodeRabbit says it still has to determine how many of the extra findings are genuine vulnerabilities. More findings is not the same as more true findings.
How much does GPT-6 Cyber cost?
The model ID is gpt-6-cyber. Prices below come from the Cyber models table on the OpenAI pricing page, checked on November 6, 2026. Rupee figures use USD/INR 94.49, the opening rate that day per CNBC-TV18, and will drift.
| Per 1M tokens | GPT-6 Cyber (USD) | GPT-6 Cyber (INR) | GPT-5.6 Cyber (USD) |
|---|---|---|---|
| Input | $20.00 | about ₹1,890 | $12.50 |
| Output | $100.00 | about ₹9,449 | $75.00 |
| Cached input | $2.00 | about ₹189 | $1.25 |
Cache writes for GPT-6 Cyber are $25.00 (30 minute) and $50.00 (12 hour) per 1M tokens on the pricing page. No long-context rates are published. Output tokens dominate the bill in agentic exploit work, so cached input helps less than it would in a code-review pipeline.
How do you get access to GPT-6 Cyber?
Access is by approval, not by sign-up:
- Apply to OpenAI's Daybreak programme and show authorised defensive cybersecurity work.
- Once approved for Daybreak Red, request GPT-6 Cyber. Approval is staged, so you may start with GPT-5.6 Cyber.
- Allowlist your IP ranges. Default IP allowlisting applies to Daybreak Red API access.
- If you hold a Zero Data Retention agreement, expect Private Safety Processing.
- Call
gpt-6-cyberthrough the API or Codex, and follow OpenAI's recommendations on sandboxing and action review.
OpenAI also introduced Daybreak Red for Critical Infrastructure and a $1 billion Daybreak for Frontline Defenders grant programme to subsidise access for resource-constrained defenders (OpenAI). New checks block high-risk activity on critical systems, so expect some refusals even once approved.
Which model should you use?
Approved Daybreak Red teams doing exploit validation, vulnerability research or patch development should use GPT-6 Cyber, accepting the higher token price for the benchmark gains. Teams with Daybreak Red access but no GPT-6 Cyber approval yet should use GPT-5.6 Cyber in the interim.
Everyone else should stay on GPT-6 Astra, which handles code review, vulnerability triage and patch development as a mainline model. For how Astra compares with its rivals, see our GPT-6 Astra vs Claude Fable 5.1 verdict. For a smaller grounding model measured against Astra, see GroundingPi XPeng 4B vs GPT-6 Astra. If your interest is coding agents rather than security, read SWE-2 Flash vs GLM-5.3 Flash. Finders of bugs outside this programme may also want our note on the OpenAI bug bounty changes.
The limits are plain. Every headline number is self-reported, the independent evidence is a single early evaluation, and the Critical rating tells you OpenAI itself regards the capability as dangerous in the wrong hands.
FAQ
Q: What is GPT-6 Cyber?
A: GPT-6 Cyber is an OpenAI model trained to find and validate exploitable vulnerabilities while staying within the scope a researcher defines. It was announced on November 5, 2026 (OpenAI).
Q: How much does GPT-6 Cyber cost?
A: It costs $20.00 per 1M input tokens, $100.00 per 1M output tokens and $2.00 per 1M cached input tokens, according to the OpenAI pricing page. At USD/INR 94.49 (CNBC-TV18) that is about ₹1,890 input and ₹9,449 output.
Q: Can anyone use GPT-6 Cyber?
A: No. It is limited to a select group of approved Daybreak Red customers, with access set to expand progressively (OpenAI).
Q: Is GPT-6 Cyber better than GPT-6 Astra?
A: On cyber tasks, OpenAI says yes, including when Astra does not refuse. On web pentesting recall it reports 40.3% against 32.0% for Astra (OpenAI). Astra remains the right choice for general work.
Q: What does the Critical Preparedness rating mean?
A: OpenAI rates GPT-6 Cyber Critical for cybersecurity, up from High for GPT-5.6 Cyber, which is why access is gated and extra safeguards ship with it. Details are in the system card.
Q: Is there independent testing of GPT-6 Cyber?
A: One early evaluation exists. CodeRabbit reported 22 caught issues against 19 on 33 cases, and says it still has to check how many extra findings are real (CodeRabbit).
Last verified: November 6, 2026. Benchmark and pricing figures are taken from OpenAI's published materials and may change.
Top comments (0)