OpenAI Deploys GPT-6 Cyber With 48.5% Exploit Rate and $100 Output Pricing

(SAN FRANCISCO —) OpenAI has formally introduced GPT-6 Cyber, a purpose-built security variant of its flagship model architecture engineered to push the boundaries of automated vulnerability research, exploit generation, and defensive remediation. Operating under strict administrative oversight via the gated Daybreak Red enterprise access tier, the model arrives with a significantly elevated cost structure — billing $20 per million input tokens and $100 per million output tokens — substantially outpacing competing enterprise cyber models from Anthropic and Google.

Built as a specialized derivative of GPT-6 Astra, the new cybersecurity engine incorporates highly tuned reinforcement learning objectives designed to reward successful program crash reproduction and exploit construction while strictly penalizing unauthorized boundary violations. Its arrival marks a pivotal escalation in autonomous software security, driving performance metrics on complex benchmarks well past previous-generation ceilings while enforcing rigid enterprise compliance frameworks.

Technical Architecture & Engineering Breakdown

GPT-6 Cyber represents an architectural pivot from general-purpose reasoning into targeted, domain-specific autonomous task execution. Built directly upon the foundational backbone of GPT-6 Astra, the model incorporates specialized safety-aligned reinforcement learning (RL) loops. Unlike standard base models that optimize purely for next-token prediction or generic conversational utility, GPT-6 Cyber's reward infrastructure explicitly optimizes for complex, multi-step validation loops: recognizing memory corruption signatures, isolating instruction pointer deviations, and writing functional proof-of-concept exploit scripts.

Under the hood, the model's system card details an explicit dual-objective constraint mechanism. The first objective drives the model to maximize success metrics in reproducing crashes and constructing functional exploits. The second, competing objective heavily constrains the model's state space to strictly respect user-defined operational boundaries, authorization parameters, and systematic reporting accuracy. This delicate optimization balance is what allows GPT-6 Cyber to achieve a Critical designation under OpenAI’s Preparedness Framework without violating its operational guardrails.

Benchmark Evaluation & Rigorous Comparison

To quantify the model's capabilities, OpenAI subjected GPT-6 Cyber to a battery of rigorous internal and standardized evaluations. On ExploitGym — a benchmark designed to test whether an AI agent can successfully target and exploit a specified software vulnerability — GPT-6 Cyber achieved a score of 48.5%. This marks a substantial 16.4-percentage-point absolute improvement (approximately 51% relative gain) over its predecessor, GPT-5.6 Cyber, which scored 32.1%.

In web application testing conducted entirely without access to application source code, GPT-6 Cyber autonomously discovered 40.3% of injected vulnerabilities, outperforming base Astra models sitting at 32%. Furthermore, across RemoteExploitBench, which evaluates models against known CVEs in enterprise software frameworks including GitLab, PostgreSQL, and Jenkins, GPT-6 Cyber successfully completed 95% of assigned tasks. Interestingly, Astra models constrained under standard Daybreak Blue parameters refused 100% of these tasks, highlighting the critical role that specialized fine-tuning and access configuration play in dictating agentic utility.

Deployment, VRAM Footprint & Inference Optimization

Access to GPT-6 Cyber is heavily restricted to vetted enterprise security teams through OpenAI’s gated Daybreak Red program. Unlike general commercial endpoints, applicants must submit explicit documentation outlining their defensive workflows, defensive scopes, and intended telemetry usage before receiving API tokens. OpenAI has enrolled upward of 2,000 entities into its broader security programs, but access to the Cyber tier remains tightly throttled.

From an economic standpoint, the model's pricing reflects its high compute overhead and specialized training regimen. OpenAI lists GPT-6 Cyber at $20 per million input tokens, $100 per million output tokens, and $2 per million cached input tokens. These rates surpass competing offerings from Anthropic and Google, placing a heavy financial burden on automated continuous security pipelines. However, empirical telemetry demonstrates that the investment yields tangible results: security researchers utilizing the model successfully reproduced 1,724 Linux kernel vulnerabilities and formulated functional exploits for 948 of them — a dramatic leap over the 166 exploits generated by GPT-5.6 Sol in identical testing environments.

# Authenticate and query GPT-6 Cyber via Daybreak Red API endpoint
export OPENAI_API_KEY="sk-proj-****************************************"

python3 -c '
import openai
client = openai.OpenAI()

response = client.chat.completions.create(
    model="gpt-6-cyber",
    messages=[
        {"role": "system", "content": "Analyze target binary and formulate reproduction harness."},
        {"role": "user", "content": "Target: linux_kernel_v6.12_netfilter_cve. Target constraints: authorized testing harness active."}
    ],
    max_tokens=4096,
    temperature=0.2
)
print(response.choices[0].message.content)
'