GPT-5.5 as Scientific Co-Researcher: Ramsey Proofs, Gene Expression, and Cyber Defense

The Ramsey Number Proof: AI at the Boundary of Human Knowledge

Advanced mathematical reasoning has historically been one of the hardest categories for large language models because free-form proof sketches are not enough; the work must survive formal verification. GPT-5.5 challenges that limitation in a concrete way. OpenAI says an internal version of GPT-5.5 with a custom harness helped discover a new proof about off-diagonal Ramsey numbers [1].

Ramsey theory studies the conditions under which ordered structure must emerge within sufficiently large combinatorial systems. OpenAI describes the result as a proof of a longstanding asymptotic fact about off-diagonal Ramsey numbers that was later verified in Lean. That distinction matters: the claim is not simply that a model wrote a persuasive explanation, but that the mathematical argument passed machine-checkable verification [1].

“[The model] autonomously gathers evidence, tests experimental biological assumptions, and draws novel scientific conclusions.”

Professor Derya Unutmaz, Jackson Laboratory immunologist, describing GPT-5.5 Pro on 28,000-gene datasets [1]

The philosophical implication extends beyond a single result. A model that can contribute to Lean-verified mathematics is no longer just a summarization tool for researchers; it becomes part of an iterative research loop where conjecture, search, implementation, and formal checking reinforce each other [1].

Bioinformatics and the 28,000-Gene Frontier

In biological sciences, GPT-5.5’s capabilities are being validated at the frontier of real research rather than constructed evaluation scenarios. Leading immunologists at the Jackson Laboratory, including Professor Derya Unutmaz, are actively deploying the model to interpret gene-expression datasets encompassing nearly 28,000 individual genes — datasets of a complexity and scale that would require teams of specialists working for extended periods to analyze manually. The model autonomously gathers supporting evidence, tests experimental biological assumptions against available data, and draws scientific conclusions that Unutmaz reports are genuinely novel [2].

On BixBench, a standardized evaluation for real-world bioinformatics and biomedical data analysis, GPT-5.5 scored 80.5% — a meaningful 6.5-point improvement over GPT-5.4’s 74.0%. On GeneBench, a multi-stage evaluation targeting complex genetics and quantitative biology workflows, the base model scored 25.0% and the Pro variant reached 33.2%. The GeneBench absolute scores appear low in isolation but are representative of the genuine difficulty of multi-stage quantitative biology reasoning, where current human expert completion rates on the same tasks are themselves far from perfect [1][3].

Cybersecurity: 81.8% on CyberGym and the Governance High-Wire

On CyberGym, OpenAI reports GPT-5.5 at 81.8%, ahead of GPT-5.4 at 79.0% and Claude Opus 4.7 at 73.1%. The result is dual-use by nature: the same planning and debugging capabilities that help defenders find and patch vulnerabilities can lower the barrier for misuse if access is not governed carefully [1][2].

OpenAI’s Preparedness Framework treats GPT-5.5’s biological/chemical and cybersecurity capabilities as High. In the release post and system card, OpenAI says the model went through targeted red-teaming for advanced cybersecurity and biology, testing with external experts, and nearly 200 early-access partner workflows before release [1][2].

For general access, OpenAI describes stricter classifiers for potential cyber risk, stronger controls around higher-risk activity, and protections against repeated misuse. In parallel, Trusted Access for Cyber provides verified defenders with expanded access to cyber-permissive models for legitimate security work under trust and security requirements [1][5].