For three years the clinical AI race had one scoreboard: speed and quality of answers. That race produced remarkable products and is now, in an important sense, over, because the leaders have converged near the same standard. OpenEvidence's expansion into education, CME from real questions, society teaching materials in the workflow, and now EvidenceGrade rating the certainty of evidence behind each answer, is the clearest signal of what replaces it. The competition is no longer who answers fastest. It is who helps clinicians learn fastest.
The signal, read closely
OpenEvidence did not need education features to win usage; by its own reporting it was already the most widely used medical AI platform in the US, with a valuation in January 2026 of 12 billion dollars. That it invested anyway, building CME in April 2025, integrating educational content from societies such as ACEP, partnering with Cochrane in March 2026, and shipping GRADE-based evidence grading in July 2026, tells you where a dominant player believes durable value lives. Note that UK clinicians cannot currently verify any of this first-hand: the platform withdrew from the UK and EU in April 2026, a gap we analysed here. The signal still travels even where the product does not.
The same move, everywhere
Scan the rest of the field and the pattern repeats. AMBOSS, which began in medical education, runs AI Mode Learning as a study copilot beside its clinical mode. GPnotebook attached automatic CPD capture to its AI Answers at launch. UpToDate moved CME into the answer workflow. Heidi built Evidence to put cited guidance beside its scribe. And the frontier labs are circling the same territory from above: Microsoft and Google have both published research systems aimed at clinical diagnostic reasoning, work whose obvious eventual application is training clinicians as much as advising them. When incumbents, challengers and platform companies all make the same turn, the turn is structural.
Why learning is the harder, better game
Answering is a converging capability: given similar retrieval over similar corpora, systems drift toward similar answers, and switching costs stay low. Learning is different on every axis. It requires longitudinal state, a model of what this clinician knows and forgets. It produces compounding value, because a platform that measurably improves you is one you do not leave. It plugs into mandatory, budgeted professional structures, CME in the US, CPD and appraisal in the UK. And it is harder to fake, because education that does not involve effortful retrieval eventually reveals itself in outcomes. The moat, in short, moved from the answer to the learner.
What good looks like in the new race
The risk of any gold rush is theatre, and education theatre is easy: log every answer read, call it learning, print a certificate. The platforms that will deserve to win the new race are the ones building the unglamorous machinery of real learning: retrieval practice rather than exposure, spacing rather than cramming, misconception diagnosis rather than generic explanation, and transparency about evidence quality of exactly the kind EvidenceGrade gestures toward. Clinicians choosing tools should apply that filter ruthlessly.
What it means for buyers and educators
The shift changes evaluation criteria for anyone who purchases or commissions these tools. When platforms competed on answers, assessment meant accuracy audits and source checks. When they compete on learning, the questions become educational: does usage measurably improve unaided performance, does the credit awarded correspond to anything a regulator would recognise as development, and does the system's engagement design serve retention or merely retention of the subscriber? Medical educators, meanwhile, gain leverage they should use: the platforms now want what faculties have, curricula, assessment expertise and accreditation standing, which is an opportunity to shape products around learning science rather than retrofitting complaints afterwards. The organisations that engage early will get tools built to their standards; those that wait will get engagement mechanics with a certificate printer.
The scoreboard for the new race
If the competition really has moved from answering to learning, the evidence that settles it will move too. Answer quality was measured with benchmarks and accuracy audits; learning quality will be measured, eventually, in harder currency: published trials showing users of a platform improve on unaided performance, credit and credentials that regulators and royal colleges formally recognise, and longitudinal data demonstrating retention rather than momentary engagement. No platform, OpenEvidence included, has yet published that class of evidence for its education layer, which means the new race is genuinely open and marketing currently outruns proof everywhere. Clinicians can hold the whole category to a simple standard: the winner of an education race should be able to show that its users learned. Watch which companies start trying to show exactly that, because they are the ones taking the shift seriously.
iatroX and the shift
This shift is one iatroX anticipated rather than reacted to: the platform has paired a UK-guideline-grounded answer engine with adaptive question banks, spaced repetition, reflective CPD and a Socratic Tutor from the outset, on the thesis that answering and learning belong in one system. Watching the largest company in the category arrive at the same conclusion is, frankly, encouraging. We track the strategic landscape continuously for UK clinicians and organisations.