Most AI companies pitch themselves as general-purpose problem solvers. Corti went the other direction. The Copenhagen-based lab spent years obsessing over a single, gnarly problem that sits at the intersection of healthcare, bureaucracy, and human error: medical coding. On April 1, the company released Symphony for Medical Coding, and the benchmarks aren't subtle. Up to 25 percent more accurate than OpenAI, Anthropic, Amazon, Oracle, and Google. That's not a rounding error. In a field where a single missed code can mean a suicide attempt goes untracked or a hospital loses tens of thousands in reimbursement, 25 percent is the difference between a system that works and one that pretends to.

Symphony is now live via API, available to any team building AI-powered healthcare software. In a market dominated by horizontal model providers trying to bolt medical features onto general architectures, Corti's vertical bet looks increasingly smart. The company, co-founded by Andreas Cleve and Lars Maaloe, has built its reputation by doing fewer things better than anyone else.

Here's a number that should bother you. In a recent study of Danish patient data, Corti found three times as many suicide attempts as human coders had recorded. The cases were all there, buried in clinical notes and medication records. Coders working under time pressure simply missed them. Three times. That's not an edge case. That's a systematic failure embedded in the infrastructure of every health system that relies on manual coding.

When cases go uncounted, health systems can't monitor trends. They can't allocate resources. They can't design interventions. Policy doesn't just fail at execution. It fails before it starts. A government that doesn't know suicide attempts tripled in a specific demographic can't respond. The data never made it into the system.

Medical coding isn't a prediction problem. It's a reasoning task. The American system alone, ICD-10-CM, has over 70,000 diagnosis codes. Guidelines evolve constantly. Models trained purely on historical data go stale fast. That's what makes Corti's approach different. Symphony doesn't just pattern-match. It reasons through hierarchies, validates against live guidelines, and reconciles ambiguity the way a professional coder would. But faster. And without the fatigue that causes human coders to miss three out of every four suicide attempts.

Symphony didn't appear out of nowhere. Corti conducted what it calls the largest study of its kind, analyzing 5.8 million patient encounters. That research produced Code Like Humans, a multi-agent framework accepted at EMNLP 2025, one of machine learning's top conferences. The framework mirrors the steps a professional coder takes: identify evidence in clinical notes, reason through code hierarchies, validate against current guidelines, and reconcile the inevitable ambiguity when multiple codes could apply.

Symphony builds on this foundation, turning academic research into a production-ready API. It's a transition that kills most academic AI projects. The gap between a research paper and a product that handles real patient data under real regulatory constraints is enormous. Most teams don't survive the crossing.

Corti's team includes researchers who've published at NeurIPS, ICML, and now EMNLP. That academic pedigree matters because medical coding sits at the boundary of natural language understanding, structured reasoning, and domain expertise. You can't fake competence here. The codes either map correctly to clinical reality or they don't, and the downstream consequences of getting it wrong are measured in patient outcomes, not user engagement metrics.

It's a playbook that should make horizontal AI providers nervous. While OpenAI and Anthropic build models that do everything adequately, Corti built one that does a single thing exceptionally well. In healthcare, where adequacy can mean malpractice, that distinction matters enormously.

There's a running debate in AI investing about whether vertical or horizontal models will capture more value long-term. Corti just handed the vertical camp a powerful data point.

The argument for vertical AI is deceptively simple. General-purpose models optimize for breadth. Vertical models optimize for depth. In domains where depth determines whether your output is useful or dangerous, the vertical model wins. Medical coding is the perfect test case because the consequences of 'good enough' are measured in denied claims, missed diagnoses, and regulatory violations.

Cleve has been making this argument for years. Medical coding isn't something you solve by throwing a bigger model at it. It requires domain-specific reasoning that general models simply don't have. The regulatory landscape shifts constantly. The coding guidelines are dense, contradictory, and region-specific. Symphony handles both the American ICD-10-CM system and the European equivalent, which means it's navigating two completely different regulatory frameworks simultaneously. Balderton Capital backed Corti's $60M Series B in 2023. The company operates across the US and Europe, with its engineering core still rooted in Copenhagen.

Corti doesn't behave like a typical AI startup. No breathless launch events. No claims about artificial general intelligence. Just a press release, an API endpoint, and benchmark numbers that speak for themselves. That restraint is itself unusual. In an industry where most companies announce partnerships before they've shipped a product, Corti shipped the product and let independent benchmarks do the talking.

The company's headquarters sit in Copenhagen's growing tech district, a few blocks from where Novo Nordisk and Denmark's biotech cluster have established the country as a serious player in health innovation. Corti fits that ecosystem perfectly: deeply technical, clinically rigorous, and allergic to hype.

For healthcare systems drowning in coding backlogs, the pitch is straightforward. Plug in the API. Get better results than you'd get from OpenAI or Anthropic. Don't wait for a general-purpose model to catch up to what a specialized one already does.

Symphony's launch arrives at a moment when hospitals and insurers are desperate for AI that actually works in production. Not demos. Not pilots that never graduate. Production systems that handle real patient data under real regulatory constraints. The number of health-AI pilots that have died in procurement purgatory is staggering. Most hospitals move at geological pace.

Corti's bet is that the companies building production healthcare systems won't want to fine-tune a general model themselves. They'll want a purpose-built API from a team that's spent years on the specific problem. It's the picks-and-shovels play for healthcare AI, and if the benchmarks hold up in real-world deployments, it could become the standard infrastructure layer for medical coding globally.

The big question isn't whether Symphony works. The benchmarks suggest it does. The question is whether healthcare buyers, notoriously slow and risk-averse, will move fast enough to matter before a larger player catches up. Corti's done the hard engineering. Now comes the arguably harder challenge: navigating procurement cycles that can stretch longer than some startups survive. In healthcare AI, patience isn't just a virtue. It's a survival skill.