OpenAI GPT-6 Cyber launches for approved organizations only
OpenAI has launched GPT-6 Cyber to support vulnerability research and fix validation. It is initially available to Daybreak Red organizations with model-specific approval. Whether organizations based in South Korea can obtain approval or use it has not been disclosed.
GPT-6 Cyber is initially available only to Daybreak Red organizations that have received separate approval for the model.
Initial access is through the API and Codex.
Standard short-context pricing is $20.00 for input and $100.00 for output per 1 million text tokens.
OpenAI reported an ExploitGym score of 48.5% for GPT-6 Cyber.
Whether organizations based in South Korea can obtain approval or use it, and when access will be fully expanded to all Daybreak Red customers, have not been disclosed.
OpenAI announced GPT-6 Cyber on November 5, 2026. Initial access is limited to Daybreak Red organizations with separate model approval. It is available through the API and Codex.
OpenAI's announcement, documentation, and system card were reviewed as of November 7, 2026.
Model launched to support vulnerability research and fix validation
GPT-6 Cyber supports vulnerability research and validation of fixes. The release was listed in the OpenAI API changelog. Reproducing vulnerabilities requires an authorized test environment.
OpenAI described the following uses:
· Vulnerability research
· Reproducing vulnerabilities in authorized test environments
· Validating fixes
Separate model approval also required for API and Codex access
Initially, only Daybreak Red organizations approved for GPT-6 Cyber can use it. OpenAI's announcement specified the API and Codex as the initial access channels. Joining Daybreak Red alone does not grant access to the model.
Item checked | Details stated in official sources
Initial eligible users | Daybreak Red organizations with separate approval for GPT-6 Cyber
Access channels | API and Codex
Access for organizations based in Korea | Approval and availability not disclosed
Expansion to all Daybreak Red customers | Completion date not disclosed
API access requires an IP allowlist by default. This setting specifies which IP addresses are allowed to connect. OpenAI's documentation explains that Daybreak-related controls are initially OFF. Approval alone does not automatically turn on refusal mitigation.
Standard short-context output pricing is $100 per 1 million tokens
The standard short-context output price in the model documentation is $100.00 per 1 million tokens. Pricing is based on text tokens. All amounts below are in US dollars.
Billing item | Standard short-context price per 1 million text tokens
Input | $20.00
Cached input | $2.00
Cache writes | $25.00
Output | $100.00
The context window specified in the model documentation is 1,050,000 tokens. The maximum output is listed as 128,000 tokens. The context window and maximum output are separate limits.
The context window indicates how much context the model can process. Maximum output refers to the limit on the output it can generate. When reviewing costs, check the conditions under which standard short-context pricing applies. The conditions for cache writes should also be checked separately in the model documentation. (Verified figures: standard short-context cache writes, $25 per 1 million tokens, 30-minute retention, $50, 12-hour retention · Source: developers.openai.com · Checked 2026-11-07)
ExploitGym score of 48.5% in OpenAI's evaluation
OpenAI reported an ExploitGym score of 48.5% for GPT-6 Cyber. The GPT-5.6 Cyber score in the same announcement was 32.1%. The company described the gap as about 16 percentage points.
Model | ExploitGym score | Results reported by
GPT-6 Cyber | 48.5% | OpenAI
GPT-5.6 Cyber | 32.1% | OpenAI
The system card separately describes the ExploitGym Honeypot evaluation. With deployment safety instructions applied, the observed rate of exploitation of out-of-scope services was 0%. This is the result OpenAI reported under those evaluation conditions.
OpenAI also disclosed capability classifications under its Preparedness Framework. The cybersecurity classification is Critical. Biological and chemical capabilities were rated High. These classifications and the ExploitGym score are separate evaluation items.
Availability in Korea and exact announcement time not disclosed
Official sources did not specify whether organizations based in Korea can receive approval or access. The announcement date is confirmed, but the time and time zone were not disclosed. The announcement's date and time in Korea therefore cannot be determined.
Undisclosed item | Information that cannot currently be confirmed
Organizations based in Korea | Eligibility for GPT-6 Cyber approval and access
Announcement timing | Exact time and time zone, date and time in Korea
Expansion of access | Completion date for expansion to all Daybreak Red customers
Independent measurements | Artificial Analysis measurement results for GPT-6 Cyber
OpenAI's evaluation results should be identified as company-reported figures. The official sources covered in this article contain no independent measurement results from Artificial Analysis. There is no basis for labeling these as independently verified performance figures.
Korean organizations should first check approval scope and billing conditions
Korean organizations should first check whether they can obtain separate approval for GPT-6 Cyber. The model documentation directs users to confirm access with their OpenAI account representative. Availability for organizations based in Korea should not be assumed to be a publicly established fact.
· Confirm with your OpenAI account representative whether organizations based in Korea can receive approval and access.
· Check whether you have GPT-6 Cyber model approval separate from Daybreak Red approval.
· If using the API, check the IP allowlist configuration required by default.
· Check the status of Daybreak-related controls. Approval alone does not turn on refusal mitigation.
· Before estimating costs, check the conditions under which short-context pricing and cache billing apply.
Vulnerabilities must be reproduced in authorized test environments, as specified in the API changelog. The system card recommends carrying out security work in isolated environments. It also recommends supervision that allows a human to stop high-risk tasks.
Injoys
