Cortex Engine Benchmarks
These benchmarks score Cortex against de-identified primary-care charts with real coding and documentation gaps. The useful question is not which model sounds smartest. It is which engine catches more clinically relevant issues without becoming slow, expensive, or noisy.
Claude Sonnet 5 — now running in Cortex — is right about 56% of what it flags, versus 37% for the previous model on the same review pipeline, at matching overall accuracy (46% vs 45%). Fewer wrong flags means fewer interruptions you learn to ignore — that is the trade we optimize for, and why it was promoted on 2026-07-01.
Claude Fable 5 led the five new models at 50% overall accuracy and 58% trust-weighted accuracy. Its range still overlaps Cortex today, while a median chart took 22.9s versus 15.5s. That is not enough evidence to trade away speed for an uncertain gain, so Cortex stays on Claude Sonnet 5.
Every candidate runs the same 10 de-identified real charts and is scored by an independent judge model. Promotion candidates are repeated three times to check that the result is stable. A model reaches Cortex only after it improves the current balance, especially by avoiding false alarms without missing too much.
10 charts is a small sample, so every score is published with its range. When two models' ranges overlap, we call it a tie instead of declaring a winner. Speed and cost are published too (1.1s per chart at the fast end, 0.11¢ at the cheapest) — the numbers vendors usually leave out.
GPT-OSS 120B (Cerebras)
OpenAI GPT-5.4 Nano
Claude Fable 5
Claude Fable 5
Model comparison
Start with Cortex today and four useful reference points. Add any evaluated model when you need a deeper comparison.
5 of 15 models selected
Production always includedCircle size shows median cost per chart (smaller is better).
Open a row for precision, recall, and test details.
| Model | Overall | Trust-weighted | Time | Cost | Readout | Details |
|---|---|---|---|---|---|---|
1 Cortex today · Claude Sonnet 5 In Cortex today | 46% | 53% | 15.5s | 2.5¢ | The model reviewing charts now | |
2 Claude Fable 5 Latest candidate | 50% | 58% | 22.9s | 10.0¢ | Best new result; not a clear enough gain to switch | |
3 OpenAI GPT-5.6 Sol Latest candidate | 42% | 54% | 14.7s | 3.6¢ | Very precise and responsive, but missed too many issues | |
4 GPT-OSS 120B (Cerebras) Comparison | 21% | 28% | 1.1s | 0.16¢ | Fastest tested, misses more | |
5 OpenAI GPT-5.4 Nano Comparison | 29% | 34% | 5.9s | 0.11¢ | Cheapest tested |
The model reviewing charts now
Scoped scoring: graded on specificity, missing-dx, audit-risk, em-level because this configuration is not designed to produce every benchmark category.
Best new result; not a clear enough gain to switch
Very precise and responsive, but missed too many issues
Fastest tested, misses more
Cheapest tested
Benchmark method
Every candidate reviews the same 10 de-identified primary-care charts with real coding and documentation mistakes. Synthetic smoke tests are never included in the published scores.
An independent judge creates the expected issue list and checks each result. We report precision, recall, overall accuracy, time, and cost. Cortex never grades its own work.
Candidates run three times. Score ranges show how much the result could move, and overlapping ranges count as a tie. This is still a small, primary-care-only test set, so it guides product decisions. It is not a clinical validation claim.