Overview
OpenAI on August 10, 2026 launched GPT-5.6-Cyber, a large language model (LLM) purpose-trained for advanced, authorized cybersecurity work. <cite index="6-4,6-5">OpenAI announced the new cybersecurity-focused model alongside an expansion of its Daybreak Cyber Partner program.</cite> <cite index="6-6,6-7">Built on GPT-5.6 Sol, GPT-5.6-Cyber has been trained for specialized tasks such as finding zero-days and creating exploit chains, and is configured with a lower refusal rate for risky dual-use tasks.</cite>
Access Architecture: Daybreak Blue and Red
<cite index="18-13,18-14">OpenAI expanded its Daybreak cybersecurity program with two access tiers: Daybreak Blue, which focuses on defensive tasks such as malware analysis, and Daybreak Red, which is geared toward offensive security research.</cite> <cite index="26-2,26-3">In production, OpenAI deploys system-level safeguards to screen cybersecurity-related requests to prevent misuse, but those filters can also block legitimate defensive work; Daybreak Blue access removes those guardrails.</cite> <cite index="26-4,26-5,26-6">Even without system-level guardrails, GPT-5.6 Sol will still refuse certain highly dual-use prompts, such as penetration testing of production systems; GPT-5.6-Cyber, available through Daybreak Red, was trained to further reduce those refusals and improve performance on such tasks for trusted defenders.</cite>
Performance and Benchmarks
<cite index="19-5">OpenAI's internal Advanced Cybersecurity Completion Rate benchmark—which measures willingness to assist with exploit-chain development, authentication bypass, and privilege escalation tasks—shows GPT-5.6-Cyber completing 95% of such requests.</cite> <cite index="22-3,22-4">GPT-5.6 Sol completes just 1.5% of those requests, and 2.0% when used with Daybreak Blue access; GPT-5.5-Cyber, the predecessor model, finished 57.3%.</cite> <cite index="23-8">Beyond benchmarks, OpenAI said it used GPT-5.6-Cyber to uncover two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome, that could be chained to corrupt memory and escape the V8 heap sandbox.</cite>
<cite index="24-7,24-8,24-9">The Advanced Cybersecurity Completion Rate measures only whether a model responds to a request—not whether the response is accurate or produces a working result—and the wider benchmark picture is described as mixed.</cite> <cite index="19-1">Under OpenAI's Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as reaching the "High" cybersecurity capability threshold but remained below the "Critical" threshold.</cite>
Pricing and Access Controls
<cite index="20-2">OpenAI's documents list pricing for GPT-5.6-Cyber at $12.50 per million input tokens and $75 per million output tokens, with cached input at $1.25 per million tokens.</cite> <cite index="7-3,7-4">Access is limited to select companies including Accenture, IBM, Capgemini, Cognizant, EY, KPMG, PwC, NCC Group, and SpecterOps, and is rolling out to supported security vendors including Palo Alto Networks, CrowdStrike, Cisco, Sophos, Akamai, Fortinet, and Cloudflare.</cite> <cite index="7-5,7-6">OpenAI says it will not give regular users access to the underlying models, citing security risks; approved partners will instead deploy them inside existing security products, managed services, and customer engagements.</cite> <cite index="19-7">To mitigate misuse risks tied to reduced safeguards, OpenAI is mandating hardware security keys for all individual Daybreak accounts starting September 1, 2026, and rolling out enhanced monitoring in the coming weeks.</cite>
Competitive Context
<cite index="14-5">According to AI Release Tracker, a new OpenAI model arrives approximately every 37 days.</cite> The GPT-5.6-Cyber launch comes amid an intensifying frontier model race. <cite index="29-15,29-16,29-17">The summer of 2026 has shaped up as the most compressed frontier release window in AI history, with OpenAI, Anthropic, xAI, DeepSeek, and Z.ai all moving simultaneously—the first time the top model position has been genuinely contested across five labs in the same calendar quarter.</cite> <cite index="32-1">DeepSeek has intensified a pricing competition with U.S. artificial intelligence providers that it first triggered in January 2025 with its R1 model, which claimed frontier-level reasoning performance at a fraction of the cost of comparable OpenAI products.</cite> <cite index="8-2">The GPT-5.6-Cyber launch comes just days after OpenAI said it was delaying the release of its forthcoming model, Astra, after it reached critical hacking abilities during safety testing.</cite>