Corti, a maker of AI-based models for healthcare, has released a new agent model for medical coding that outperforms many Big Tech models.
Symphony for Medicalcoding outperforms OpenAI, Anthropic, Amazon, Oracle, and Google benchmarks by more than 25% on clinical accuracy benchmarks. The product, available via API, is built on Corti’s flagship model Symphony, which is already used by 200 teams in the US today. The company works with EHR vendors, virtual care platforms, practice management systems, and life sciences organizations around the world.
Medical coding is extremely complex. The U.S. coding system, ICD-10, has approximately 70,000 diagnosis codes. Automated processes transform clinical notes into structured data to inform reimbursement, research, and policy.
It is very important to be able to catch the right chord within the nuances. Errors can be costly both in terms of lost revenue and lack of diagnosis. In a recent study of Danish patient data, Corti found three times as many suicide attempts were coded. Overlooking these trends has implications for resource allocation and intervention design.
“Most AI systems fall short for medical coding because they treat medical coding as labeling rather than inference. Correct coding depends on interpretation of evidence, context, hierarchy, and guidelines,” Dr. Lars Marloe, co-founder and chief technology officer of Corti, said in the announcement. “We built Symphony for Medicalcoding to follow the same decision-making process that professional programmers use, which is why the performance gap is so meaningful.”
According to Corti, Symphony was evaluated on two public benchmarks widely used in medical coding research: ACI-BENCH and MDACE. They were created by an independent academic team and expert medical programmers. Symphony for Medicalcoding has also been validated with real-world clinical data from a large U.S. health system across emergency and ambulatory care settings. All models compared, including Corti and Big Tech, were evaluated under identical conditions and each was run five times to ensure consistency and reproducibility.
Specifically, Corti’s model outperformed Anthropic’s Claude Opus 4.6, OpenAI’s GPT-5.4, and Google’s Gemini 3.1 Pro. It was also compared to systems built on these foundational models, including Oracle Health & AI’s MedDCR (built on GPT) and AWS AI’s fine-tuned coding model (built on Claude).
According to Corti, existing models that automate coding typically rely on memorizing patterns from annotated datasets through supervised or semi-supervised learning. This approach is not optimal for rare code, multiple specialties, or frequent system updates. Corti’s model is based on targeted LLM and reimagines medical coding as a reasoning process using four agents that mimic the work of human programmers. The company says its agents identify evidence, reason through hierarchies, verify it against guidelines, and reconcile ambiguities.
With the advent of ambient scribes, unstructured clinical notes are becoming longer and the need for clear coding is greater than ever, Marloe told Fierce Healthcare. This means you’re more likely to miss opportunities to write code. Corti’s primary use case is revenue cycle management, but Maaløe says the platform can also track signs of fraud.
Corti complies with HIPAA and GDPR privacy standards. Maaløe has discovered that US customers are looking for Corti products that are GDPR-level compliant, as the US lags behind privacy regulations.
Corti was founded in 2016 as a research company seeking to understand how large-scale language models can be streamed in real-time to reduce administrative burden and enhance clinician workflow. It has raised $100 million so far.
Steve West, managing director of Healthliant Ventures and Tanner Health, said in the announcement that Corti’s methodology is “the most promising approach to medical coding we’ve seen to date. We’ve been co-developing it with Corti because we believe specialized AI infrastructure is the way to solve this problem, and we look forward to seeing it go into production.”
