AI's Next Race Is Cost per Answer, Not Model Size, and Medicine Will Feel It

Most AI models write the way a typist does, one word-piece at a time. Two young companies are betting that the next contest in AI is about doing the same job for far less money. If they are right, the price of putting AI into every clinic and phone falls fast.

Our research desk led its September 27 edition with two stories. The first is Inception, co-founded by Stanford computer scientist Stefano Ermon. Its Mercury models use diffusion, the method behind many image generators, to draft many tokens at once and refine them over a few steps. Ermon argues that this uses graphics chips more fully, and that the next era of competition will turn on efficiency, not raw size.

The second is Jev, from a startup called TypeSafe. Jev does not write prose at all. It answers structured questions: yes or no, pick from a list, place on a scale. The desk reported claims that Jev runs about 200 times faster and 400 times cheaper than general models on those jobs, with early users compressing a million-token conversation, speeding a computer-use agent, and sorting thousands of emails for pennies.

Inception and Ermon. Ermon did appear on the No Priors podcast in September 2026 under the theme that diffusion could win AI inference. The episode summary we found matches the desk's account in broad terms: parallel generation, better hardware use on standard GPUs, efficiency over size. That summary was built from the episode title and description, so we treat the details as Ermon's view, not tested fact.

Inception's own materials describe Mercury as the first commercially available diffusion language model, launched in February 2025. The company claims up to ten times the speed of comparable models and one-fifth the latency of Claude 4.5 Haiku, priced at $0.25 per million input tokens. An earlier TWIML podcast described Mercury 2 as five to ten times faster than small frontier models and pitched it at voice and agent work. None of these figures come from an independent benchmark.

TypeSafe and Jev. The New Stack confirms the basics. TypeSafe emerged from stealth with $40 million in seed funding led by DCVC. Chief executive Diogo Almeida announced Jev on September 15. His launch post claimed 20 to 200 times faster responses and 40 to 400 times lower cost. A later analysis put the headline multiples at about 194 times faster and 445 times cheaper, but against baselines TypeSafe chose and ran through its own wrapper.

The customer anecdotes are thinner than the desk implied. One developer reported shrinking a session of nearly a million tokens to about 86,000 in around a second. We found no measured result for the computer-use agent and no source for the emailsorting figure. An independent tester found Jev unreliable at counting and dates, and LangChain's outside check used only five test cases.

Medicine is full of small, repeated decisions. Is this lab value urgent? Which inbox message needs a nurse today? Does this note mention a fall? Today those questions often go to large, slow, costly models, or to no model at all. A cheap, calibrated decision engine could triage messages, flag risk and route referrals at a cost a small practice can bear.

Lower cost also changes who gets AI. When an answer costs a fraction of a cent, tools can run continuously on wearables and home monitors, not only in big health systems. That is the real promise of the efficiency race.

• Every speed and cost multiple we found comes from the companies or their chosen comparisons. • A model that cannot explain itself is hard to audit. In clinical use, a yes or no without reasons needs careful validation. • TypeSafe's benchmark measures agreement with labels made by two frontier models, not ground truth. • Faster is not the same as safer. Error rates on real clinical tasks have not been published.

I would welcome a cheap, fast model that sorts my inbox by urgency. I would not trust one until it had been tested on real patient messages against a nurse's judgment, with misses counted honestly. Cost per answer matters. Cost per missed emergency matters more.

• Independent benchmarks of diffusion language models on accuracy, not only speed. • Peer-reviewed or third-party tests of Jev-style decision models on labelled realworld data. • The first health-system pilots that report cost per correct triage decision.

Claim What We Found Tier

Your Wearable's Heart Rate Variability May Be Partly Sleep Apnea. A Testable Idea, Rated 45 Percent AI's Next Race Is Cost per Answer, Not Model Size, and Medicine Will Feel It Staggered Shifts Leave More People Unaware of a Rumour, Our Simulation Suggests