OpenAI releases GPT-5.5

OpenAI released GPT-5.5, an update to its GPT-5 line pitched around agentic reliability rather than a single headline benchmark: the company said the model needed less step-by-step guidance, made more efficient tool calls, recovered better from errors mid-task, and stayed coherent over long-running jobs lasting minutes to hours. President Greg Brockman called it “a faster, sharper thinker for fewer tokens” than GPT-5.4 and framed it as a step toward a unified “super app” combining ChatGPT, Codex and an OpenAI browser. It launched across ChatGPT’s Plus, Pro, Business and Enterprise tiers.

The accompanying system card described the model going through OpenAI’s Preparedness Framework evaluations, including red-teaming for cybersecurity and biological-risk capability, using feedback from close to 200 early-access partners. OpenAI reported that sustained jailbreaking by expert red-teamers had produced biological-safety failures at the model level during predeployment testing — a finding that shaped the safeguards shipped alongside the model rather than one that stopped release.

Alongside launch, OpenAI opened its GPT-5.5 Bio Bug Bounty, inviting vetted security and biosecurity researchers to find a single “universal jailbreak” prompt able to make the model answer all five questions in an internal biosafety challenge from a clean session without tripping moderation, restricted to the model running in Codex Desktop. The reward was later raised to $50,000, shared with a parallel programme for GPT-5.6, with GPT-5.5 testing scheduled to run into late July 2026. Publishing both a capability jump and a live financial incentive to break its own biosafety controls made GPT-5.5 one of OpenAI’s more explicit acknowledgements that frontier capability and biological-misuse risk were rising together.