Photo: Tim Witzdam / Pexels
The two frontier models share a 41.7% pass rate on a benchmark that tests whether AI can actually digest long medical records
The two best AI models at reading long medical records now share first place. Their winning score is 41.7%.
Artificial Analysis released an update to the MLCR-AA leaderboard on October 21, 2026. The update left Anthropic’s Claude Sonnet 5.5 and OpenAI’s GPT-6 Astra tied at the top, each with a 41.7% pass rate.
The benchmark comes from Wisedocs, an insurtech platform focused on AI-assisted review of medical and insurance records. It is designed to measure something insurers care about: whether AI can do the record synthesis work claims professionals handle today.
What the leaderboard actually shows #
The median score across models on MLCR-AA remains below 15%. The leaders are clearing a bar most of the field cannot reach at all. No model on the board breaks past 42%.
The scoring also splits into more specific measures, and the leaders differ depending on which one you check. Claude Sonnet 5.5 tops the completeness metric at 64.8%. OpenAI’s GPT-6.1 Sol leads on accuracy at 96.7%.
Those two figures tell different stories. Accuracy asks whether what the model says is right, while completeness asks whether it caught everything it should have.
The leaderboard data points to Anthropic models performing particularly well on completeness. OpenAI’s lineup holds the edge on accuracy.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Why long context is the hard part #
The benchmark feeds models synthetic medical and insurance case files containing 25,000 to 64,000 tokens. Because the case files are synthetic, the benchmark can test realistic record review without exposing real patient data.
What changed in the update #
The October 21 update incorporated practical reviewer prompts and refined 15 of the benchmark’s 60 expert references. Those references are the expert-written answers that model outputs get graded against.
Before the update, earlier models like Claude Fable 5 had posted a 64.4% completeness score. Comparing that figure directly with post-update numbers is not straightforward given the revised answer key.
Background: a benchmark built for claims work #
Wisedocs launched the original MLCR benchmark in June 2026. Its stated goal was to measure how well AI systems can synthesize medical data in the way claims professionals do.
MLCR-AA is the version of the evaluation hosted on Artificial Analysis.
What this means for insurers and AI vendors #
A top pass rate of 41.7% means even the best models fail the majority of these expert-level synthesis tasks. With the median below 15%, picking a model at random from the leaderboard is a poor bet.
The split between completeness and accuracy leaders gives buyers a real tradeoff to weigh. A team worried about missed information might lean toward Anthropic’s models, while a team worried about confident errors might lean toward OpenAI’s.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our