GPT-5.5 appears in the research record as a large language model evaluated alongside Claude Opus 4.7 in a retrieval-augmented study on the Nuclear Cardiology Board Preparation Exam, where both systems were reported to surpass average human performance. The work was first posted as a medRxiv preprint (doi 10.64898/2026.05.08.26352768v1) and subsequently appeared in the Journal of Nuclear Cardiology as article 102946, available online on 18 August 2026, published by Elsevier. The study, led by Killekar, Shanbhag, Miller, Dey, and colleagues at Cedars-Sinai Medical Center, frames GPT-5.5 as a model capable of absorbing retrieval-supported clinical content and answering specialist board-style questions at a level above the human reference group used in the comparison.
In that evaluation that quick-info value, GPT-5.5 demonstrates the kind of domain-grounded reasoning that fits retrieval-augmented clinical workflows, where a model integrates supplied reference material with its own knowledge to answer nuanced specialty questions. The pairing with Claude Opus 4.7 in the same study suggests GPT-5.5 is positioned by researchers as a peer-tier system for high-stakes medical knowledge tasks rather than a lightweight assistant. Beyond this single published evaluation, no first-party documentation of the model's training recipe, architecture, parameter count, or open-weight status was provided in the source material, so claims about its lineage or training scale should be treated as unverified.
