Two days ago, a company called Actava released a one trillion parameter AI model built specifically for healthcare, and almost nobody outside the industry noticed. No launch video. No breathless keynote. Just a technical announcement, a benchmark table, and a price list. Which is strange, because the price list might be the most interesting AI story of the month.
The model is called CURA 1T. I want to walk through what it actually is, why building for hospitals instead of consumers might be the smarter play, and what the cost claims mean if they hold up.
What Actava actually shipped
CURA 1T is a one trillion parameter model built on top of Kimi-K2.6, an open foundation model, and then pushed further with recursive self-improvement, which is a fancy way of saying the model helped generate and refine its own training pipeline for the healthcare domain. It handles text and images, supports function calling, and works with a 256K context window, which is enough to hold a serious chunk of a patient chart in a single request.
Actava says it posted the strongest performance on five of six healthcare benchmark panels they tested against. Benchmarks are marketing until proven otherwise, and I will get to the caveats. But the specific numbers are worth knowing because they map to real jobs.
The company positions the model around three use cases.
First, patient-facing care and triage. On HealthBench Professional, a benchmark built from real clinician-graded conversations, CURA scored 66.2, which Actava says is the highest of any model tested. This is the “patient messages the portal at 11pm asking if their symptom is an emergency” problem. Health systems drown in these messages, and nurses spend hours a day answering questions that follow predictable patterns. A model that handles the routine 80 percent and escalates the scary 20 percent is not science fiction. It is a staffing problem with a software answer.
Second, clinical reasoning. The model was trained across 17 medical specialties and 11 body systems. On HealthBench Hard, the deliberately brutal version of the benchmark, it scored 36.8, a 14.6 point jump over its base model. That number sounds low until you realize the test is designed so that most models score near the floor. The point is not that the model is a doctor. The point is that the healthcare-specific training moved the needle a lot, which tells you the domain tuning is real and not just a new coat of paint on a general model.
Third, and this is the one I find most interesting, agentic workflows inside electronic health records. On MedAgentBench, a test of whether a model can actually operate against FHIR, the standard interface hospitals use to move patient data around, Actava reports CURA completed tasks with a 94 percent success rate. Think ordering labs, filing referrals, pulling records, and updating charts. This is the unglamorous plumbing of medicine, and it is where clinicians burn hours every day. If a model can reliably drive these systems, you are not talking about replacing doctors. You are talking about deleting paperwork.
The cost angle, and why it matters more than the benchmarks
Here is the number that made me sit up. Actava claims CURA delivers its results at 5 to 20 times lower cost per output token than the frontier models it outperforms on these panels. Actava’s stated pricing is 50 cents per million input tokens and 2.50 dollars per million output tokens, served through an OpenAI-compatible API.
Why does that matter? Because healthcare AI is a volume business. A single hospital system generates millions of patient messages, chart summaries, and administrative tasks a year. At frontier-model prices, running AI across all of that is a budget line that makes a CFO sweat. At these prices, it starts to look like a rounding error next to the labor it offsets. The economics decide whether this technology gets deployed at the pilot-project scale or the whole-health-system scale, and pricing at this level is a bet on the second.
The OpenAI-compatible API detail matters too. It means a hospital’s software vendor can test CURA by changing a URL, not by rebuilding their stack. That combination is how challengers take share from incumbents.
And this is why I think enterprise healthcare is a smarter wedge than consumer AI right now. Consumer is a knife fight over subscriptions with the biggest companies on earth. Healthcare enterprise is a market where the buyer has a painfully specific problem, a budget already allocated to solving it, and a workflow where “good enough plus cheap plus reliable” beats “brilliant but expensive” almost every time.
The honest take
Now the cold water. Actava itself says CURA is a research model and not a substitute for a clinician. Take that seriously, because they clearly do. Benchmarks are curated environments. Real hospitals are chaos: incomplete charts, contradictory records, patients who describe chest pain as “a weird feeling.” The gap between a 94 percent benchmark score and a system you would trust with your mother’s referral is filled with validation studies, liability lawyers, and compliance reviews that take years, not quarters.
I also cannot independently verify the benchmark claims or the cost comparisons. Those are Actava’s numbers about Actava’s model. Every vendor’s launch chart shows their bar as the tallest one. The MedAgentBench figure deserves particular scrutiny: 94 percent task success is roughly 25 points higher than independently published state-of-the-art scores on that benchmark, which makes it an extraordinary claim by any standard. The right posture is interested skepticism: the claims are specific and testable, the pricing is stated publicly, and the API is open enough that third parties will check the math soon.
But here is what I keep coming back to. Even if CURA is only half as good as claimed, a specialized trillion-parameter model at commodity prices, wired into the actual data standards hospitals run on, is a preview of where this market goes. The future of AI in medicine probably is not one genius model that does everything. It is purpose-built models doing specific, boring, expensive jobs at prices that make deployment a no-brainer.
If you run a business, in healthcare or anywhere else, the lesson travels: the AI opportunity in your industry is probably not the flashy demo. It is the paperwork.
What is the most expensive boring task in your business? That is the question worth sitting with this week. If you want more plain-English breakdowns of where AI is actually moving the needle, follow along.
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com