| ||||||||||||||||||||||||||||||||||||||||||||||||
“In every high-stakes decision – in justice, in healthcare, in policing – machines can inform, but humans must decide – and answer.” António Guterres, Secretary-General of the United Nations Opening the first UN Global Dialogue on AI Governance · Geneva · July 6, 2026 | ||||||||||||||||||||||||||||||||||||||||||||||||
| The Model Race | ||||||||||||||||||||||||||||||||||||||||||||||||
By the Numbers · AI Model and Capability Tracker
Artificial Analysis Intelligence Index v4.1, higher is better (highest-scoring configuration per model). Context in tokens; input/output in USD per million tokens, standard non-cached first-party rates. Open-weight models can be self-hosted, so hosted prices vary; Inkling has no first-party per-token rate (n/a). *Kimi K3 open weights release July 27. Sources: Artificial Analysis, OpenRouter, and vendor pricing pages. Claude is made by Anthropic. | ||||||||||||||||||||||||||||||||||||||||||||||||
The AI race is tilting from frontier labs to open models, which enterprises favor for cost, control, and ownership. Chinese open-weight models now top download charts, and half the Fortune 500 run open or private deployments. Moonshot AI’s Kimi K3, the largest open model yet at 2.8 trillion parameters, matches frontier coding at $15 per million output tokens, far below closed US rivals. For health systems, the economics of closed APIs are under pressure. BBC · TechCrunch | ||||||||||||||||||||||||||||||||||||||||||||||||
Mira Murati’s Thinking Machines released Inkling, its first model, with open weights on Hugging Face and fine-tuning via its Tinker platform. The 975-billion-parameter system activates 41 billion per task and lets enterprises customize on their own data and infrastructure, a pitch aimed at health systems wary of locked-down frontier APIs. Axios | ||||||||||||||||||||||||||||||||||||||||||||||||
DeepMind CEO Demis Hassabis argues AGI is likely only a few years away and calls for a FINRA-style frontier AI standards body, initially voluntary, to test top models for cybersecurity, bio, and deception risks before US release. He frames the moment as needing cautious optimism over a race. Demis Hassabis | ||||||||||||||||||||||||||||||||||||||||||||||||
| Deals and Dollars | ||||||||||||||||||||||||||||||||||||||||||||||||
UnitedHealth and Elevance both beat Q2 estimates and raised 2026 guidance, signaling a managed-care rebound. UnitedHealth posted adjusted EPS of $6.38 against a $4.90 estimate and lifted its full-year outlook to $19.50 to $20.00 on tighter medical cost control. Elevance guided to $27 EPS and will exit more Medicaid markets. Reuters (UNH) · Reuters (Elevance) Q2 2026 Scorecard · UnitedHealth vs Elevance
Company Q2 2026 reports (UnitedHealth July 16, Elevance July 15). Adjusted EPS is not directly comparable across companies. Market cap and P/E as of mid-July 2026 and will move with the market. P/E shown is trailing; UnitedHealth’s is elevated by weak prior-year earnings, so its forward P/E is roughly 22x versus about 11x for Elevance. | ||||||||||||||||||||||||||||||||||||||||||||||||
Optum, part of UnitedHealth Group, said July 13 it will partner with Anthropic, deploying Claude to cut administrative burden while keeping human oversight of clinical judgment. The company disclosed no specific use cases or timeline. The deal follows Optum’s $3 billion AI push and ongoing litigation over AI-driven coverage denials. Becker’s Payer | ||||||||||||||||||||||||||||||||||||||||||||||||
Neko Health, the Stockholm preventive-scan startup cofounded by Spotify’s Daniel Ek, raised $700 million in Series C funding, pushing total backing past $1 billion. It will open its first US clinic in New York this year, bringing its AI-driven full-body imaging and blood-test model stateside amid debate over whole-body screening. Fierce Healthcare | ||||||||||||||||||||||||||||||||||||||||||||||||
| AI in Practice | ||||||||||||||||||||||||||||||||||||||||||||||||
Mayo Clinic is partnering with Scale AI on clinical tools that streamline pre-visit chart review, flag safety events like wrong-site surgeries and falls, and help staff work at top of license. Physicians have gained an average of 11 more minutes per patient, per Scale AI’s own figures. Becker’s Hospital Review | ||||||||||||||||||||||||||||||||||||||||||||||||
Harvard, Dana-Farber, and Mass General Brigham researchers built ALADYNOULLI, an interpretable AI model that predicts a patient’s risk across 348 diseases using EHR data and genetics. Tested on 683,000 people, it tracks how risk evolves over decades and outperformed standard calculators for cardiovascular and breast cancer prediction. Published in Nature. Inside Precision Medicine | ||||||||||||||||||||||||||||||||||||||||||||||||
CVS Health unveiled Health100, an AI assistant built with Google Cloud that schedules checkups, fills prescriptions, and checks coverage. CEO David Joyner wants it to become healthcare’s default AI front door across payers and providers. A small test group comes first, with wider rollout later this year amid antitrust scrutiny. Axios | ||||||||||||||||||||||||||||||||||||||||||||||||
The AMA is advancing CPT codes for autonomous AI that generates clinical outputs without a physician present. Nursing leaders are wary, echoing the ANA’s warning that AI must not replace professional judgment and raising accountability and patient-safety concerns about a physician-controlled coding system governing autonomous tools. Becker’s Hospital Review | ||||||||||||||||||||||||||||||||||||||||||||||||
How was this issue?
Or just reply to this email. We read every response. |
AI in Healthcare | What Comes Next Ep. 18
As open models narrow the capability gap and reshape enterprise economics, healthcare is entering a more contested AI era, one defined by governance standards, human accountability, payer recovery, and the race to turn intelligence into measurable clinical and operational value.
Jul 20, 2026
•
3 min read