GPT-5.5
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models
- Lineage
- Neurons and networks
- Kind
- Implementation
Who
OpenAI
What happened
OpenAI releases GPT-5.5, focused on agentic coding, computer use, and office work. The company rates its cybersecurity capabilities as "high" and ships it with stricter safeguards.
Why it mattered
Better at long agentic tasks. It ships with tightened controls because of its cybersecurity capability.
The full story
On April 23, 2026, seven weeks after GPT-5.4, OpenAI released GPT-5.5 and presented it as a new class of intelligence for real work. Its strengths were agentic coding, computer use, office work, and the early stages of scientific research.
According to OpenAI, it scored 82.7 percent on Terminal-Bench 2.0, against 75.1 percent for GPT-5.4, and 78.7 percent on OSWorld-Verified, and it used fewer tokens than its predecessor on the same tasks. The company also said the model had helped discover a new proof about Ramsey numbers, later verified with Lean, a program that checks proofs.
Cybersecurity shaped the launch, which came two weeks after Claude Mythos Preview. Under its Preparedness Framework, OpenAI rated the model's capabilities in cybersecurity and in biology and chemistry as "high," short of the "critical" level. It shipped the model with stricter classifiers and offered fewer restrictions to verified defenders through a trusted access program. On May 7 it announced GPT-5.5-Cyber, a limited-access variant for cybersecurity teams.
It went first to ChatGPT subscribers, to developers the next day, and on May 5, in its Instant version, to free users. The AI Security Institute measured a 71.4 percent success rate on expert-level cybersecurity tasks, above Claude Mythos Preview's 68.6 percent.
Leads to
GPT-6 Astra (September 3, 2026)
Documented direct influence
OpenAI rates both models under its Preparedness Framework, GPT-5.5 at the high level for cybersecurity and GPT-6 Astra at the critical level, and Astra's FrontierMath score is set against GPT-5.5's
Threads that pass through here
- The model that was too capable (step 2 of 5)
Sources
Scholarship and context
Further reading
How to cite
“GPT-5.5.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/gpt-5-5
This entry was last reviewed on November 13, 2026.