General-purpose models beat the specialized clinical AI hospitals already pay for

verifiedHealthcare

Monday, June 15, 2026

Conviction

High

Time horizon

This quarter

Risk

Paying a vertical-tool premium for accuracy a frontier model now matches

The News

A study in Nature Medicine found general-purpose frontier models — GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6 — outperformed the specialized clinical tools OpenEvidence and UpToDate Expert AI across MedQA, HealthBench, and a benchmark built from real physician queries. Source: Nature Medicine.

The Read

The vertical-AI pitch in regulated fields rests on an assumption this punctures: that a domain-tuned tool is inherently safer or more accurate than a general model. It cuts directly against the Salesforce thesis two stories up — Salesforce paid $3.6B betting defensibility lives in a post-trained vertical model, while in clinical Q&A the general models won outright. The reconciliation is that post-training buys workflow integration and liability ownership, not necessarily raw accuracy, and buyers should stop paying a premium for the latter when what they're really getting is the former. Before the next clinical-decision-support renewal, the frontier model is the baseline your specialized vendor has to beat — not merely clear.

Counterview

Benchmark Q&A is not bedside use; a specialized tool's value is its curated, auditable, guideline-linked sourcing, which a higher MedQA score doesn't capture and a malpractice review would still demand.

Do This Week

Pull your top 50 real clinical-reference queries and run them blind against both your current specialized tool and a frontier model, scoring accuracy with a clinician. Walk that delta into the renewal conversation — if the vertical tool doesn't clear the frontier baseline, you're paying for sourcing and workflow, so negotiate on that, not on accuracy claims.

For Healthcare