GPT-5.5 System Card: OpenAI's Smartest Model Gets Its Most Rigorous Safety Review Yet

OpenAI has published the System Card for GPT-5.5, their newest and most capable model, detailing the extensive safety evaluations and red-teaming conducted before release.

GPT-5.5 represents a significant leap over GPT-5.4, scoring 82.7% on Terminal-Bench 2.0 (vs 75.1%) and 39.6% on FrontierMath Tier 4 — nearly doubling Claude Opus 4.7's 22.9%. On knowledge work benchmarks (GDPval), it achieved 84.9%, outpacing Anthropic's best by 4.6 points.

Notably, the model independently discovered a novel proof pathway for off-diagonal Ramsey numbers, verified through formal proof tools like Lean — marking genuine mathematical contribution rather than mere reproduction.

On the safety front, OpenAI launched a dedicated Bio Bug Bounty offering up to $25,000 for universal jailbreaks targeting biological safety risks. The company describes GPT-5.5 as carrying "the strongest set of safeguards to date."

Key specs: 400K context window, $5/$30 per million input/output tokens (2x GPT-5.4 pricing), co-designed with NVIDIA's GB200/GB300 NVL72 infrastructure. An NVIDIA engineer noted that losing access to GPT-5.5 would feel "like amputation" — highlighting how deeply integrated it has become in professional workflows.

The model uses fewer tokens to complete equivalent tasks while maintaining the same latency, breaking the typical trade-off between capability and speed. Currently available through Codex and rolling out to paid ChatGPT subscribers, with API access coming soon.