OpenAI has begun a restricted preview of its GPT-5.6 model family, introducing three capability tiers named Sol, Terra, and Luna. The company describes GPT-5.6 Sol as its most capable model for cybersecurity work, emphasizing its strength in defensive tasks such as vulnerability identification and patch development.

Tiered Architecture and Branding Shift

The release marks a deliberate change in OpenAI’s versioning approach. Going forward, version numbers denote the model generation while names represent fixed capability tiers. GPT-5.6 Sol is positioned as the flagship for high-intensity reasoning. GPT-5.6 Terra targets everyday workloads and reportedly matches GPT-5.5 performance at roughly half the operating cost. GPT-5.6 Luna occupies the lowest pricing tier, optimized for speed.

Cybersecurity Performance and Limitations

In evaluations on the ExploitBench benchmark, Sol matched competing systems while consuming approximately one-third the output tokens. Testing against the Chromium and Firefox codebases showed the model could isolate bugs and basic exploitation primitives, but it did not independently construct a working full-chain exploit. OpenAI characterized Sol’s orientation as leaning toward defensive security rather than end-to-end offensive operations.

Multi-Layered Risk Controls

To address dual-use concerns, the GPT-5.6 series employs several overlapping safeguards beyond standard training-level refusals:

  • Automated real-time classifiers monitor biology and cybersecurity inputs during inference.
  • When an anomaly is flagged, output generation pauses while a secondary reasoning model reviews the conversational context.
  • Some flagged interactions can trigger account-level evaluations to distinguish legitimate security research from malicious use.

OpenAI also reported investing more than 700,000 A100-equivalent GPU hours in automated red-teaming, with a focus on finding universal jailbreaks rather than single-prompt failures.

Phased Rollout and Government Coordination

The restricted launch followed consultation with the U.S. government and is aligned with a recent Executive Order directing federal assessment of national security risks posed by advanced AI systems. Current access is limited to approved partners through the API and Codex. OpenAI plans to expand availability to ChatGPT, broader API users, and Codex in the coming weeks.

The company has publicly stated its opposition to making government-mediated pre-clearance a permanent requirement, arguing that prolonged access restrictions delay critical defensive tools from reaching the wider security community.