GPT-5.6-Cyber
Who gets in, who doesn't
OpenAI split Daybreak into Blue and Red tiers, and the offense-oriented GPT-5.6-Cyber is offered only under Red — application-vetted, unavailable through ordinary API channels.
Four key numbers
Advanced security task completion
OpenAI's internal advanced cybersecurity benchmark. The previous GPT-5.5-Cyber scored 57.3%; unguardrailed Sol manages only ~2%.
Per million tokens, input / output
Compiled by third-party pricing trackers; the official blog post doesn't list prices. Roughly 2.5x GPT-5.6 Sol ($5 / $30).
V8 0days found pre-release
Two previously unknown Chrome V8 bugs that chain into a heap-sandbox escape; one is CVE-2026-15903, already patched by Google.
TAC individual compliance deadline
Individual members must complete Advanced Account Security with a hardware passkey before 2026-09-01 or lose access to high-capability models.
What the Daybreak split is
On 2026-08-10 OpenAI published "Expanding Daybreak as the Cyber Defense Window Narrows", splitting its cyber-defense program Daybreak into two tiers: Blue offers GPT-5.6 Sol without system-level cybersecurity guardrails, for defensive work like vulnerability discovery, secure code review and malware analysis; Red opens the dedicated offense-capable model GPT-5.6-Cyber for authorized vulnerability research and exploit validation. Cyber is built on Sol and is the first model rated High for cybersecurity capability under OpenAI's Preparedness Framework.
What's happened since launch
The launch came with a 16-company security-vendor partner program (IBM, CrowdStrike, Accenture, Cisco, Cloudflare and others); from 08-13 both Daybreak tiers became available to qualified customers via Amazon Bedrock. Meanwhile the TAC (Trusted Access for Cyber) hardware-passkey requirement entered enforcement: introduced on 06-01 alongside the Yubico partnership announcement, individual members must comply by 09-01 or fall back to standard model access. On the regulatory side, the White House convened OpenAI, Anthropic, Google and others on 08-04 to discuss a frontier-model cybersecurity testing framework.
Timeline
The TAC hardware-passkey requirement begins with the Yubico partnership announcement, covering access to high-capability cyber models.
Daybreak splits into Blue / Red and GPT-5.6-Cyber launches, alongside disclosure of the two V8 bugs found pre-release.
Compliance deadline for TAC individual members' hardware passkeys; anyone who misses it reverts to default model access.
Confirmed vs caution
Confirmed
The release date, the two-tier structure, the 95.0% / 57.3% / ~2% completion rates, the 16 partners, the Bedrock channel and CVE-2026-15903 all come from OpenAI's official blog and help center. The $12.5 / $75 pricing is compiled by third-party pricing trackers — OpenAI didn't list prices in the post itself.
Caution
Claims like "a third-party platform already lists GPT-5.6-Cyber" or "we can activate Daybreak Red for you, no vetting" contradict the official line: Red is application-vetted, requiring identity verification, usage monitoring and a lawful-use commitment. Treat any such offer as high-risk.
Blue or Red
Daybreak Blue
Defensive. GPT-5.6 Sol without system-level cyber guardrails, covering vulnerability discovery, code review and incident response; ~2% completion on advanced security queries.
Daybreak Red
Authorized offense. The dedicated GPT-5.6-Cyber model for exploit validation and security testing, at 95.0% completion; vetted access, monitored throughout.
How to apply
The only entry point is openai.com/daybreak/partners: submit organization and use-case details, pass identity verification, and sign the monitoring and lawful-use terms to enter the Red tier. TAC members must additionally enable a hardware passkey before 09-01. There is no self-serve API path.
On QCode
QCode does not offer GPT-5.6-Cyber. The verifiable facts on OpenAI's side: the model is supplied only in the Daybreak Red tier, which requires an application and review, and TAC members must additionally enable a hardware passkey. For defensive security work (code review, vulnerability analysis), start with GPT-5.6 Sol, which is in stock: official pricing multiplied by our service rate, pay-as-you-go.
FAQ
Can ordinary API access GPT-5.6-Cyber?
No. Per OpenAI's own documentation, it is supplied only in the application-vetted Daybreak Red tier, requiring an application, identity verification and monitoring terms. QCode does not carry this model; treat any "we can get you access" offer with caution.
What's the difference between Blue and Red?
Blue gives defensive work the unguardrailed GPT-5.6 Sol; Red gives authorized offensive research the dedicated GPT-5.6-Cyber. Both require an application; Red's vetting and monitoring are stricter.
What does GPT-5.6-Cyber cost?
Third-party pricing trackers put it at $12.5 input / $75 output per million tokens; OpenAI's official blog post doesn't list pricing, so cite it with that caveat.
What vulnerabilities did it find?
During pre-release evaluation it found two previously unknown Chrome V8 bugs that chain into a heap-sandbox escape; one is CVE-2026-15903, already patched by Google (Chrome versions before 150.0 are affected).
What's the TAC hardware passkey about?
TAC members must complete Advanced Account Security with a hardware passkey. The requirement started 2026-06-01; the individual-member deadline is 2026-09-01, after which access to high-capability cyber models is lost.
Which model for defensive security work?
Daybreak Blue, or the general-purpose GPT-5.6 Sol. Sol is available on QCode ($5 / $30 per million tokens), billed at official pricing times our service rate — good for getting started with code review and vulnerability analysis and for day-to-day volume.
Sources
OpenAI official blog (2026-08-10), OpenAI help center TAC entry, Yubico official blog (2026-06-01), Help Net Security (2026-08-11), research notes from Forrester and the Cloud Security Alliance; pricing compiled by third-party pricing trackers.
Defensive security work starts with a flagship in stock
Seven model families including GPT-5.6 Sol are callable with one key on QCode, billed at official pricing times our service rate.
Related reading
GPT-5.6 family guide
How to pick between Sol / Terra / Luna.
Sol vs Terra vs Luna
Capability and pricing across the three same-generation tiers.
MCP security sandboxing
Isolation and boundary practices for tool calls.
Not affiliated with OpenAI. This page compiles public information and offers no access to any restricted tier.