OpenAI released GPT-6 Cyber on November 5, 2026 with a 1,050,000-token context window and a 128,000-token output limit. It suits approved Daybreak Red security teams doing authorized exploit development, while everyone else should pick GPT-6 Astra for code review and patching.
Can you use it?
- OpenAI API: gated (Organizations cleared by the program team). First wave: a select group; OpenAI promises wider access soon.
- Codex: gated (Cleared program members using Codex). Same first wave as the API.
- Free plan: not published.
- Gate: Daybreak Red, Apply through Daybreak; IP allowlisting is on by default for API access.
Checked 7 Nov against OpenAI's own pages. Compare access across frontier models
What changed
GPT-6 Cyber is OpenAI's security research model, announced November 5, 2026. OpenAI reports 48.5% on ExploitGym, up from 32.1% for GPT-5.6-Cyber, and rates it Critical in cybersecurity under its Preparedness Framework. Access is limited to approved Daybreak Red customers through the API and Codex.
Where it sits
- $40.00/M$ per 1M tokensBlended price (3:1)Lower is better#94 / 95peer median $1.69/Mvendor price, checked by Hokai
Pricier than 99% of the 95 GA models with a published price, and one of 37 that document a zero-data-retention option. Ranked against GA models; this record is not GA.
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT-6
Context window: 1,050,000 tokens · Max output: 128,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About GPT-6 Cyber
OpenAI announced GPT-6 Cyber on November 5, 2026 as its most capable security research model, the successor to GPT-5.6-Cyber inside the GPT-6 generation that began with GPT-6 Astra. It is built to find and validate exploitable vulnerabilities while staying inside the scope a researcher sets. OpenAI has not disclosed its architecture, parameter count or training cutoff, so none are stated here. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for the previous Cyber model.
OpenAI's own tests show the gains in four places. On ExploitGym, which asks a model to exploit a specified vulnerability in a remote target, it scores 48.5%, up from 32.1% for GPT-5.6-Cyber. On a new offensive web penetration test where the model probes applications without source code, it reached 40.3% recall of vulnerabilities against 32.0% for GPT-6 Astra. On RemoteExploitBench, built from known flaws in GitLab, PostgreSQL and Jenkins, it completed 95% of tasks, and its one miss was an exploit it could not get to run remotely. On ExploitGym Honeypot, a scope test, it avoided attacking out-of-scope services in 99.93% of runs without a deployment prompt. All of these are OpenAI-reported figures.
OpenAI also describes field results. Its researchers used the model to develop working exploits for 948 of 1,724 reproduced Linux kernel vulnerabilities, against 166 for GPT-5.6 Sol in the same test setup, and to produce evidence for 81 novel Firefox vulnerabilities that Mozilla rated high severity. Security teams at Cisco, Calif and CodeRabbit are named as users.
The model takes text and image input and returns text. The API page lists a 1,050,000-token context window, 128,000 maximum output tokens, reasoning token support, streaming, function calling and structured outputs. In the Responses API it supports web search, file search, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. Audio, video and fine-tuning are not supported.
Access is the main constraint. GPT-6 Cyber is available through the API and Codex to a select group of approved Daybreak Red customers, with wider access to Daybreak Red customers promised soon. Daybreak Red API access requires an IP allowlist by default, and API customers on Zero Data Retention generally have to use Private Safety Processing, which keeps encrypted records in customer storage. Anyone outside the program should use GPT-6 Astra or GPT-6.1 Sol for code review and patching, or compare other security-tuned models such as Gemini 3.8 Flash Cyber.
Pricing
Standard short-context rates per 1M tokens: $20.00 input, $2.00 cached input, $100.00 output. Long-context rates were not shown on the page checked. Only approved Daybreak Red organizations can use the model.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.600 | $0.100 | $0.700 |
| Support reply | $0.040 | $0.030 | $0.070 |
| One coding agent run | $4.00 | $2.00 | $6.00 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Exploit development and validation: Built to find vulnerabilities and prove them with working exploits, with a 51% relative ExploitGym gain per OpenAI.
- Scope-aware behavior: Trained to stay within authorized scope; OpenAI measured near-perfect compliance on a honeypot test run without a deployment prompt.
- Long context window: Holds large codebases in one session and writes long outputs.
- Full Responses API tool set: Hosted shell, apply patch, computer use, MCP, web search and code interpreter work with the model.
Pros
- Largest reported ExploitGym gain between two Cyber models, per OpenAI.
- Misalignment monitoring and IP allowlisting are built into the access program.
- Real-world results OpenAI cites include 81 Firefox findings and 948 Linux kernel exploits.
Cons
- Closed program: no self-serve signup, and the price list applies only after approval.
- Costs $20 and $100 per 1M tokens, well above mainline models.
- Benchmarks come from OpenAI; independent replication is not yet published.
Frequently Asked Questions
What does GPT-6 Cyber cost per million tokens?
OpenAI's API page lists $20.00 per 1M input tokens, $2.00 per 1M cached input tokens and $100.00 per 1M output tokens at the standard short-context rate. Cache writes cost $25 per 1M tokens for 30 minutes or $50 per 1M for 12 hours. Input costs about 60% more than the $12.50 its predecessor charges, and only approved Daybreak Red organizations can buy it.
Is GPT-6 Cyber better than GPT-6 Astra at security work?
OpenAI says GPT-6 Cyber beats GPT-6 Astra on most cyber tasks, even where Astra does not refuse. On a new web penetration test it found 40.3% of vulnerabilities against 32.0% for Astra. On RemoteExploitBench, Astra with Daybreak Blue refused every task while GPT-6 Cyber completed 95%. These are vendor figures, and no independent run is public yet.
Can I run GPT-6 Cyber outside the OpenAI API?
It is proprietary and OpenAI does not release weights. Per OpenAI, early access sits with vetted program members, by API and in Codex, and a wider rollout is promised. No cloud marketplace listing for this model was found on the dates checked, though its predecessor is on Amazon Bedrock.
Does OpenAI train on what I send GPT-6 Cyber?
OpenAI's announcement says standard retention is the default. API customers on Zero Data Retention generally must use Private Safety Processing, which reviews encrypted records kept in customer storage for 30 days without exposing prompts to OpenAI staff. Training use of Daybreak Red inputs is not addressed in that post, so check your program terms.
Who should apply for GPT-6 Cyber and who should skip it?
Apply if you run authorized exploit development, large-scale vulnerability research or red-team work and can accept IP allowlisting and monitoring. Skip it if you mainly need code review, triage or patch drafting, since OpenAI points those jobs to mainline models like GPT-6 Astra. Teams that cannot get Daybreak Red approval have no purchase route at all.