Cortex Benchmarks
When a major AI company releases a new model, we run it through the same review Cortex gives your charts, on the same de-identified real charts, and have an independent judge grade the results. Cortex switches models only when a new one proves better, and above all when it raises fewer false alarms.
When it flags something, it is right 100% of the time, and it catches 89% of the issues within Cortex's checks that a careful coder would raise. Few false alarms with most issues caught is the balance we protect: a wrong flag costs your attention, so it counts double in the trust score.
The strongest result came from Claude Opus 5.5: a 95% trust score against 98% for Claude Sonnet 5.5 — about as accurate, but more false alarms than today. Cortex stays on the model it runs today until a newcomer clearly beats it.
It has to beat today's model on the same charts, with at least 9-in-10 confidence, while keeping false alarms within 2 points of today's, missing at most 10 points more issues, staying about as fast, and costing no more than twice as much. A model that loses nothing, catching as much with no more false alarms, can also earn a switch by being clearly faster or cheaper to run. A model from another company also has to be approved for patient data before it can run in Cortex.
10 charts holding 9 issues within Cortex's checks is a small sample, so every score comes with a likely range, and models are compared chart by chart instead of on a single number. Speed and cost are in the comparison below. This guides product decisions; it is not a clinical validation claim.
Claude Sonnet 5.5
Claude Opus 5.5
OpenAI GPT-6 Luna
Model comparison
Start with Cortex today and four useful reference points. Add any tested model when you want a deeper comparison.
5 of 13 models selected
Production always includedUp is more accurate, left is faster, smaller circles cost less.
Open a row for the likely range and test details.
| Model | Trust score | Right when it flags | Issues caught | Time | Cost | Readout | Details |
|---|---|---|---|---|---|---|---|
1 Claude Sonnet 5.5 In Cortex today | 98% | 100% | 89% | 2.9s | 0.66¢ | The model reviewing charts now | |
2 Claude Opus 5.5 Compared | 95% | 96% | 93% | 6.4s | 2.3¢ | About as accurate, but more false alarms than today | |
3 OpenAI GPT-6 Luna Compared | 91% | 100% | 67% | 2.0s | 0.01¢ | Less accurate than today | |
4 OpenAI GPT-6 Astra Also the judge | 98% | 100% | 89% | 3.0s | 0.61¢ | About as accurate as today | |
5 Google Gemini 3.1 Pro Compared | 95% | 100% | 78% | 13.8s | 2.4¢ | About as accurate, but misses more issues than today |
The model reviewing charts now
About as accurate, but more false alarms than today
Less accurate than today
About as accurate as today
About as accurate, but misses more issues than today
Benchmark method
Each model runs through the exact pipeline Cortex uses in the clinic, with the coding rules and the AI review working together, not a simplified test version.
Every model reviews the same 10 de-identified primary-care charts with real coding and documentation gaps, 3 times each. Made-up test charts are never part of the published scores.
An independent AI judge from a different company writes the list of issues a careful coder would raise, marks which ones fall within what Cortex is built to check, then grades every model's findings against that list. A correct catch outside Cortex's scope counts neither for nor against a model. Cortex never grades its own work, and the trust score counts a false alarm twice as heavily as a missed issue. The judge's own model is tested too; because it may favor itself, it is marked and never shown as the strongest result.
This is a small, primary-care-only test set. Today's charts test problem-list and medication review; checks that rely on the visit note or lab results are not yet part of the published score. Results guide product decisions and are not a clinical validation claim.