Medicine often has no clean “right” answer. Across 15 surgeries with multiple treatment options, physician identity explained anywhere from 7% to 77% of the variation in which option was chosen. A model can disagree with the recorded decision and still be clinically reasonable.

Benchmarks usually see less context than a doctor. The median real patient record in Protege’s data contains about 8,500 tokens. Five of six public healthcare AI benchmarks provide less context than that, and several use fewer than 200 tokens.

Small evaluation choices can change the winner. Reordering the same 19 diagnostic answer choices caused models to change answers. In another test, adding one sentence to the prompt erased performance gaps and even flipped which model came out on top.

AI is already moving into real care faster than our evals are improving.
Protege’s data suggests AI is on track to write nearly 1 in 3 SOAP notes, while patient references to AI in clinical conversations are also rising quickly.

The real scoreboard is health. Mortality, QALYs, clinician burnout, denials, medical debt, trust, and whether AI improves outcomes at the same or lower cost are ultimately what matter. Benchmark scores are useful, but they are only a proxy.

Neurosurgeons perform first live AI-assisted brain surgery: A real-time AI system watched the surgical video feed and helped neurosurgeons identify hidden arteries and optic nerves while removing a brain tumor, with the patient recovering vision within days. (link)

Datavant launches digital medical record exchange for providers: Provider Exchange uses Datavant's network of 80,000+ provider sites to let organizations securely request, track, and receive medical records digitally, reducing manual retrieval work and delays. (link)

Turquoise Health connects healthcare pricing data directly to AI: A new MCP connector lets users query Clear Rates and payer-provider pricing data from tools including Claude, ChatGPT, and Gemini. (link)

Transcarent launches bundled payments for advanced cancer therapies: The new model covers BiTE and CAR T treatments delivered through community providers, giving employers more predictable costs and potential savings of up to $150,000 per case. (link)

Tempus trains oncology foundation model on 1.67M patients: The multimodal model combines longitudinal clinical data, DNA, RNA, and pathology images into a single patient representation, improving survival prediction and treatment-benefit ranking across cancer cohorts. (link)

Google unveils GlucoFM for continuous glucose monitoring: The foundation model learns from CGM data to support metabolic health tasks including diabetes risk assessment, insulin resistance, beta-cell dysfunction, and post-meal glucose prediction. (link)

Anthropic gives Claude control of laboratory hardware: Its new Model Hardware Standard lets AI agents operate equipment such as liquid handlers and robotic arms through standardized interfaces, with Genentech testing it to automate a common drug discovery assay. (link)

Outer Bio trains AI on living human skin: The startup's Yuna platform keeps human skin viable outside the body for four weeks while machine learning prioritizes compounds to test, creating a faster feedback loop for skincare and dermatology research. (link)

Ambience launches shared AI patient context layer: Chorus builds a continuously updated, source-linked understanding of each patient from EHR data, creating common infrastructure that can power multiple clinical AI workflows. (link)

Abridge expands clinical intelligence across 300+ health systems: The company is making its context-aware clinical decision support available to every clinician at partner health systems, regardless of whether they use Abridge for documentation, as usage has tripled in two months. (link)

Suki launches AI-powered clinical dictation: Suki Dictation is built directly into Epic and Meditech and can be deployed independently or alongside the company's ambient documentation platform, giving health systems more flexibility in how they adopt clinical AI. (link)

Luma Health connects pre-visit AI to the clinic floor: New operational AI features combine patient preparation, arrival status, conversational AI, and smart queues to help staff see what each patient still needs and streamline clinic flow. (link)

Fitbit founders launch AI health wearable for families: Luffu Link combines health sensing, voice logging, location awareness, and cellular connectivity with AI that can identify changes in health patterns and help families remotely support aging loved ones. (link)

Healthline AI surpasses 1.1M views: Fullspan Health's conversational health agent has crossed 1.1M views since launching in June and now reaches roughly 13,500 users per day, with engagement up 28% since launch. (link)

ResidencyRL trains medical AI through simulated experience: The reinforcement learning approach improved diagnostic accuracy by 7% and reduced missed clinical red flags by 31%, with clinicians preferring the trained model in 87.6% of comparisons. (link)

ECRI launches nationwide reporting for healthcare AI errors: The patient safety group is asking providers to report incorrect or unsafe AI behavior after a survey found nearly one-third of leaders had encountered misleading AI output and 9% saw an error reach a patient or affect care. (link)

Mayo AI detects pancreatic cancer before diagnosis: REDMOD analyzes standard CT scans to identify signs of pancreatic cancer months before diagnosis, outperforming radiologist review on sensitivity and potentially enabling earlier detection from imaging patients already receive. (link)

Stanford tests EHR-integrated LLM for surgical triage: Across 6,193 cases, the human-in-the-loop tool identified patients needing hospitalist co-management with 94% sensitivity and 74% specificity. Only 2 of 19 false negatives were attributed to LLM misclassification, with most errors tied to workflow or clinical criteria. (link)

Northwell builds KPI framework to measure ambient AI: The health system developed strict performance metrics around ambient documentation to quantify efficiency gains and determine whether the technology is delivering measurable operational value. (link)

Mayo advances nurse-built AI agents to proof of concept: Five autonomous agent concepts designed by nurses with Microsoft are moving into testing, targeting complex administrative workflows and clinical tasks identified directly by frontline staff. (link)

Study tests leading AI models on 20,000+ spine reports: GPT-4o, Claude-4, Qwen-3 Max, and DeepSeek-V3.1 showed high accuracy interpreting clinical radiology reports, although performance remained weaker for rare spinal conditions. (link)

Tempus wins FDA clearance for pulmonary hypertension AI: The company's third FDA-cleared cardiovascular device analyzes standard 12-lead ECGs to flag signs of pulmonary hypertension, extending Tempus' AI diagnostics into another difficult-to-detect cardiovascular condition. (link)

ROPCA wins FDA clearance for autonomous ultrasound robot: Arthur autonomously performs musculoskeletal ultrasound scans while companion AI software analyzes the images and generates reports, clearing the Danish company to enter the U.S. market. (link)

Penn Highlands Healthcare + Doximity: Penn Highlands partnered with Doximity to deploy its Clinical AI Suite across the health system, including AI-powered clinical search, ambient documentation, and secure patient communication. (link)

Heidi Health + VIDAL: Heidi Health partnered with VIDAL to embed cited drug information, prescribing guidance, and interaction data directly into Heidi Evidence. (link)

Candid Health + Flatiron Health: Candid Health partnered with Flatiron Health to integrate autonomous revenue cycle management with OncoEMR, bringing AI-powered billing automation to oncology practices using the EHR. (link)

Corti + DCAI: Corti and DCAI launched a sovereign AI infrastructure platform giving European healthcare organizations governed access to AI models with controls for data sovereignty, compliance, retention, and costs. (link)

IKS Health + Western Washington Medical Group: Western Washington Medical Group deployed IKS Health’s Scribble Select, allowing clinicians to choose between AI-only, human-validated, or clinician-reviewed ambient documentation for each visit. (link)

Oura, the smart ring maker, is reportedly preparing for a $3B IPO at a $16B valuation as early as September. (link)

Adaptyv Bio, a Swiss startup building an automated lab for agentic biology using synthetic biology, nanofluidics, and machine learning, raised $40M in Series A funding led by Highland Europe. (link)

Faro AI, a San Diego, Calif.-based developer of an AI platform that helps biopharma companies design and run clinical trials, raised $37.3M in Series B funding. Merck Global Health Innovation Fund and Section 32 led the round. (link)

Metriport, an open-source healthcare data platform powering care teams and AI agents, raised $26M in funding led by Matrix, with participation from ARTIS and Y Combinator. (link)

Arintra, a San Francisco-based developer of AI software designed to help health systems code medical records, prevent claim denials, and improve reimbursement, raised $25M in Series B funding. Define Ventures led. (link)

Transfyr, a physical AI platform that captures real-world laboratory work and turns it into machine-readable data, raised $25M in seed funding led by General Catalyst, with participation from Lux Capital, Breakout Ventures, Factory, Neo, SV Angel and others. (link)

Hike Medical, which uses back-office AI agents to automate insurance approvals and digitize clinical evaluations for custom medical devices, raised $22.5M in seed and Series A funding led by Saga Ventures. (link)

Onos Health, an SF-based clinical intelligence startup, raised $17M in Series A funding. Costanoa led, joined by Flare Capital Partners and CVS Health Ventures. (link)

Legato, an AI-powered hearing-assistance glasses company, emerged from stealth with $12M in funding from Neotribe Ventures, Listen and Village Global to bring its Legato Frames to market. (link)

OmicsBank, an SF-based clinical data infrastructure company, raised $2.3M in seed funding from Redesign Health. (link)

Sword Health + Headspace: Sword Health agreed to acquire the digital mental health company for an estimated $200M-$300M in cash, adding Headspace's therapy and mindfulness offerings to its AI-powered virtual care platform. (link)

Switchboard Health + Livara Health: Switchboard Health acquired the virtual MSK management provider, embedding its program into its AI-driven care navigation platform to help providers and health plans enroll patients. (link)

Trusted Health + ShiftOS: Trusted Health acquired ShiftOS, developer of Holly, an AI scheduling agent that builds schedules and fills call-offs, integrating it into its staffing platform for hospitals. (link)

MaineHealth defends Palantir amid nurse protests (link)

When AI stops taking notes & starts seeing patients (link)

TIME100 AI 2026: The 100 most influential people in AI (link)

Doctor responds to JAMA’s AI-only medical care piece (link)

Growth Account Executive at Stedi, an AI-enabled healthcare clearinghouse 

N/A | Remote/NYC (link)

Strategic Partnerships Associate at Nitra, an AI-native operating platform 

N/A | NYC (link)

Account Executive at Brellium, an AI clinical and compliance review

$240K - $350K | NYC (link)