| StatusGoStrategy in Motion |
AI in Healthcare · A Weekly Point of View |
|
|
In This Issue
| → | The AI market at a glance: where the money is really going |
| → | Bills up, model prices down: the week in AI economics |
| → | StatusGo Insights: why more pilots don’t create value |
| → | More answers, everywhere: but does care get better? |
|
| 01 · The AI Market, at a Glance |
The takeaway: the market is paying for the pick-and-shovel layer of AI. The companies making the hardware, memory and chips (Micron, AMD, Arm, TSMC) are far ahead of the market, while the software and cloud giants everyone assumes are winning AI (Microsoft, Alphabet, Meta) are only keeping pace, and last year’s enterprise-software favorites (Oracle, Salesforce, ServiceNow) are falling behind. Being the biggest and being this year’s winner are not the same thing. |
◀ SMALLER · UNDER $1T |
$1 TRILLION AND UP · LARGER ▶ |
|
SMALL & SURGING
Under $1T · up 20%+
|
GIANTS ON A RUN
$1T+ · up 20%+
|
|
SMALL & STEADY
Under $1T · up under 20%
|
GIANTS, HOLDING
$1T+ · up under 20%
|
|
Columns: market value (under vs. over $1T). Rows: stock up 20%+ this year vs. less. Color vs. the Dow (+7.7% YTD): green ahead, yellow even, red behind. |
As of the September 25, 2026 close (YTD from the Dec 31, 2025 close, split-adjusted, excluding dividends). TSMC and Arm use their U.S.-listed shares; Alphabet includes Google. Not investment advice. |
| 02 · Bills Up, Model Prices Down |
| 1 |
A billion-dollar bill for “smarter” coding |
|
Blue Cross Blue Shield plans say hospitals’ AI coding tools quietly added $942 million to their bills over two years, as the same care started getting coded at a higher level. The study shows a pattern, not proof, and it cannot say whether the new codes are right. But the fight it points to is real: one company’s smarter billing is another’s rising cost. If you are betting on AI here, track collections and administrative savings as two separate lines, because they do not move together. Fierce Healthcare
|
| 2 |
The government’s own AI cost warning |
|
CMS chief Mehmet Oz said the quiet part out loud: in the near term, AI will push healthcare costs up, not down, because it sharpens billing before it makes care cheaper. His answer is to lean harder on accountable care, where the reward is for spending less rather than billing more. It is a useful test for any AI project on your desk. Is it chasing more reimbursement, lower total cost, or better outcomes? Those are three different bets, and only one of them lowers what the system spends. Healthcare Dive
|
This Week’s Model Releases Prices keep falling. Here is what each release means for you. |
| MODEL |
WHAT CHANGED, AND WHY IT MATTERS |
OpenAI · Sep 22 GPT-6 Sol / Luna OpenAI |
Prices cut sharply, with Sol at $2 in / $10 out and Luna at $0.10 / $0.50 per million tokens. Why it matters: high-volume paperwork just got cheaper to automate, so re-run the business case on your busiest admin tasks. |
Anthropic · Sep 22 Claude Opus 5.5 Anthropic |
Prices down 20% to $4 / $20 per million tokens, with better scores on safety tests. Why it matters: your most complex agent workflows cost a fifth less to run, though the safety claims still need your own testing. |
Google · Sep 22 Gemini 3.8 Flash TTS Google |
New text-to-speech voices, with a lighter, cheaper option for higher volume. Why it matters: more natural machine voices for reminders and phone outreach, worth a small pilot with real patients. |
MiniMax · Sep 27 M3.1-Flash-Preview MiniMax · AlphaSignal |
A new low-cost coding model in preview, with no proven edge on quality or price yet. Why it matters: one to watch, not to switch to. Let your engineers trial it before it goes near anything real. |
|
| StatusGo Insights |
|
More AI pilots don’t create value on their own.
A pilot shows what is possible. Value takes more. Before funding another AI project, insist on five things:
| 1 | Start with a measurable business problem. Name the outcome and today’s baseline, then ask if AI is even the right tool. |
| 2 | Redesign the work. Cut the steps that should not exist before automating the rest. |
| 3 | Give one leader ownership and authority. They own the result and can change the process; IT owns the tool. |
| 4 | Count the full cost. Data prep, integration, training, review, rework, and run. Hours saved are not dollars until you say how. |
| 5 | Plan for and invest in adoption. A released tool is not a used tool. Budget training and measure real use. |
The executive question: what result are we funding, who owns it, and what must change to deliver it?
|
| 03 · More Answers, Everywhere. But Better Care? |
| 3 |
Your front door is moving off your website |
|
Zocdoc wants to be the plumbing behind every “find a doctor” search, not just its own site. Its new Care Access Network feeds live scheduling into Amazon’s health AI, Google’s Gemini, and Blue Shield of California. The shift for health systems is simple and a little uncomfortable: patients now book from wherever they happen to be asking, so a stale directory or a calendar with no open slots loses you the patient long before they reach your site. Zocdoc
|
| 4 |
Free medical AI for 100 countries, with a catch |
|
Anthropic and OpenEvidence are putting free clinical decision support in front of doctors across roughly 100 lower-income countries, and they promise to tune it to local realities. The access story is genuine and worth applauding. The catch is just as real: advice trained on wealthy-country medicine can misfire where the tests, drugs, and staffing are different. The question worth asking of any tool like this is whether its recommendations still hold up against what a clinic can actually do that day. Reuters
|
| 5 |
A mental-health test that grades the AI, not the patient |
|
OpenAI has published MentalHealthBench, a test of how AI handles mental-health conversations from everyday stress to a crisis, built with more than 80 licensed clinicians across 22 countries. Good to know it exists, but read the fine print: it scores how good the answers sound, not whether patients actually get better. If you are piloting anything in behavioral health, treat it as a first screen, and do your own testing on the parts that matter most, like when the AI hands a person off and how fast it escalates. OpenAI
|
| 6 |
Finally, a model ranking built on real work |
|
Fireworks AI put out a ranking that scores models on actual professional work, including a healthcare track built partly on 500 physician-reviewed cases from Doximity. That is a more honest yardstick than the usual leaderboards, and a better place to start a shortlist. It is still only a start. A high score earns a model a tryout, not your trust. The test that counts is whether it holds up on your workflows, at your cost, with your people reviewing the output. Fireworks AI
|
| 7 |
950 AI agents found something. Now what? |
|
Anthropic set about 950 of its Claude agents loose on biological data for 21 hours, and they surfaced an enzyme system no one had described before. Human scientists then took it to the lab, though what the system actually does is still unknown. The number worth remembering is not 950. It is how much of what AI proposes survives a real experiment. That ratio, not the headcount of agents, is what will tell you whether this speeds up discovery or just hands researchers more leads to chase. Anthropic
|
What Comes Next is a weekly point of view from StatusGo on AI in healthcare. Figures as disclosed by sources; selected stories, not exhaustive. Issue of September 29, 2026. |
|
