Seven Days After OpenAI's Clinical Benchmark Dropped, a Healthcare Specialist Topped It: Corti 60.5, ChatGPT for Clinicians 59.0
OpenAI published HealthBench Professional on May 6 to showcase GPT-5.4 with extended reasoning as the state of the art in clinical AI — beating specialist physicians given unlimited time and web access. Seven days later, Corti published results showing its Symphony system scored above everything on that benchmark, including OpenAI’s own clinical product.
The Numbers
HealthBench Professional is 525 physician-authored tasks drawn from real clinician conversations, graded against criteria written and adjudicated by three or more physicians per task. One third of the benchmark is adversarial red teaming: prompts designed to surface safety failures and probe clinical judgment limits.
| System | HealthBench Pro | Care Consult | Safety (adversarial) |
|---|---|---|---|
| Corti Symphony | 60.5 | 63.5 | 87.7 |
| ChatGPT for Clinicians | 59.0 | 51.0 | — |
| GPT-5.4 (extended high) | 48.1 | — | 30.3 |
| Specialist physicians | baseline | — | 30.0 |
The care consult category — the largest slice, closest to actual clinical decision-making — shows a 12.5-point gap between Corti and OpenAI’s purpose-built clinical product. On the adversarial safety slice, Corti (87.7) is nearly three times the score of both GPT-5.4 and human physicians (both around 30).
How It Was Built
Corti Symphony is not a new model. It is a configuration of Corti’s existing Agentic Framework — a prompt and system design built on top of available infrastructure that any developer can access. Corti’s argument: the capability OpenAI ships as a proprietary application (ChatGPT for Clinicians) is reproducible through the API, at higher accuracy, by a team that spent a week with the published benchmark methodology.
Corti has been building clinical AI since its founding — voice AI for clinicians, documentation, triage, patient interaction. The company is not a frontier lab and is not training its own foundation models. It is demonstrating that post-training configuration and domain-specific system design can outperform raw model capability on a domain-specific benchmark.
The Structural Point
OpenAI published HealthBench Professional explicitly to benchmark competitive systems, enriching the hardest examples by 3.5x and comparing against physicians with unlimited time. The benchmark was designed to be hard to beat.
Corti beat it in a week. That result doesn’t mean general clinical AI is solved — HealthBench measures specific task types and a specific evaluation methodology. But it reinforces a pattern appearing across specialized domains: the benchmark a frontier lab uses to justify a proprietary application becomes a public target, and a specialist system with domain depth reaches the top faster than the lab anticipated.
The same dynamic appeared with Mozilla and Claude Mythos (271 Firefox vulnerabilities), with DeepSeek on coding benchmarks, and now here. Frontier model capability is not the ceiling for domain performance — system design and domain knowledge are.