CMMI Is Moving Faster. But Congress Is Still Asking the Wrong Question.

In 1901, the Wright brothers’ glider produced only a third of the lift they expected. They went home to Dayton, built a wind tunnel in their bicycle shop, and tested miniature wings. The results changed their next design. Their 1902 glider performed much better and led to the powered aircraft they flew in 1903.

The value of those experiments was what the brothers did with the results. An unsuccessful wing design still helped them build a successful airplane.

That is how I think about the Center for Medicare and Medicaid Innovation’s founding mandate. The Innovation Center has a statutory obligation to design and test new payment models, and if they work, they can be scaled in Medicare, Medicaid, and/or the ACA exchange markets. Looked at this way, CMMI’s only failure from rigorous experimentation is a failure to learn. Applied seriously, that is a demanding standard: each test should leave us better able to design the next one, and useful discoveries should survive the experiment that produced them.

I watched CMMI’s first wave of models as Chief Technology Officer at HHS. I now look at them as an investor. From both perspectives, I want the Center to try more ideas, learn faster, and give people a reason to build better ways of delivering care.

The current team deserves credit for picking up the pace and pursuing de novo model types with thoughtful strategic aims. The burst of announcements at the end of 2025 included ACCESS, which pays for technology-supported chronic care, and LEAD, a new ten-year accountable care model. Other announcements addressed prevention and drug pricing. I admire the commitment to heterodox design, as in ACCESS, which the CMMI team states explicitly is priced to effectively require “deflationary” cost inputs tied to outcomes; if this works, it would be a historic achievement. This is an encouraging change in tempo, and I hope it continues. Experimentation is the job Congress created CMMI to do, and there is a very good reason for this: prior CMS authorities made such efforts unwieldy, underpowered, and even if they worked, often a bridge to nowhere.

However, a slowly building congressional debate risks pushing it in the opposite direction.

In March, the Government Accountability Office reported that CMMI had tested 70 models between 2011 and 2024, with four expanded nationwide. House Budget Chairman Jodey Arrington responded: “A 5 percent success rate isn’t innovation—it’s failure.” He also said CMMI’s mission was too important to abandon. I agree with him on the mission. I disagree with his definition of success.

For one thing, 24 of those 70 models were still being tested at GAO’s January 2025 cutoff. GAO also noted that CMS had incorporated elements of models into Medicare and used earlier work to design successors. None of that progress was included in the “5 percent success rate.”. And progress hasn’t stopped. More recently, the joint-replacement model received actuarial certification for expansion in August, another reminder that the work continues beyond the GAO’s report cutoff.

The larger problem is treating the percentage of models expanded as a verdict on experimentation. A test can tell us that a service works but the payment is wrong, that an approach helps a particular population, or that an idea should be abandoned before it costs much more. Those are useful results if we act on them. A demand for fewer unsuccessful tests can just as easily produce fewer ambitious ones.

CMMI still has a financial case to answer. CBO estimated that it increased net federal spending by $5.4 billion between 2011 and 2020, far short of the savings originally expected. Congress should press hard on that result. The claim that we learned something cannot become a standing excuse for losses. CMMI should have to show what it learned, what changed as a result, and whether the next attempt performed better. That would be useful oversight. 

So, too, would a look at the effects beyond the original payment model. Under the statutory expansion test, the CMS Chief Actuary must certify that expansion would not increase net Medicare or Medicaid spending, alongside other requirements concerning quality and benefits. Crucially, they cannot use “spillover” savings or other benefits in the commercial or employer models to offset Medicare losses in that calculation. Yet these benefits are highly relevant in terms of understanding what an experiment accomplished. As Melinda Buntin and colleagues have argued, slower health spending growth may partly reflect a broader shift toward value-based care whose effects extend beyond what individual model evaluations capture.

The Oncology Care Model is the strongest example of why we need a broader definition of success. It ran from 2016 to 2022 and cost Medicare more than it saved: reductions in medical spending were outweighed by payments to participating practices. 

But researchers publishing in the Journal of Clinical Oncology studied what happened to commercially insured and Medicare Advantage patients at those practices. Across more than 150,000 patients, they found that participation was associated with $6,287 less spending per six-month treatment episode relative to comparison practices, after accounting for baseline differences. They found no associated deterioration in the quality measures examined. In other words, savings while maintaining quality, which is just what CMMI is looking for.

What they found, and what’s crucial to accurately measuring the success of a model, is evidence of benefits outside the population Medicare was paying practices to manage. It’s intuitive — a practice that changes how it chooses drugs or coordinates care may find it easier to use the same approach for patients with different insurance. But it requires an extra step by researchers. If we want to know whether that work was worth doing, we need to follow all of its effects.

Some models might not measure up to this broader standard. The Medicare Diabetes Prevention Program shows the weakness of the expansion metric from the other direction. It cleared the expansion bar, yet CMS’s evaluation counted just 9,015 participants from April 2018 through March 2024. Suppliers described burdensome billing and payments that did not cover their costs, which CMS documented. The lesson in this failure was to fix delivery and access; just because a model is adopted doesn’t automatically get patients into the program.

Right now, Kidney Care Choices offers a current test of whether CMMI can learn and adjust. CMS reported improvements in home dialysis and certain transplant outcomes, alongside approximately $304 million in net Medicare losses in 2023. It changed the payment terms for 2026, ended one option, and extended the others through 2027. We should now ask whether the revised economics preserve those clinical gains (though recent independent analyses have sown doubt on these). The financial loss is a reason to change the experiment, while the possibility of clinical gains gives us something specific to preserve.

ACCESS makes the need for broader measurement even more obvious, and raises the stakes. Payers representing 165 million members across Medicare Advantage, Medicaid and commercial coverage have voluntarily pledged to offer arrangements aligned with its payment principles by 2028. CMS has brought a substantial part of the insurance market to the table. Tracking which contracts follow, who receives care and what happens to costs and outcomes should be foundational to evaluating the effort.

This isn’t an esoteric or conceptual debate. Today, as GP and co-founder of Evidenced LP, I work with and meet many founders and teams who are trying to build within these tailwinds from CMMI models. And I have some specific pieces of advice for those founders and the policymakers who want to conduct real, productive oversight of CMMI.

For founders, this is why the underlying need matters more than the life of any one model. A payment experiment can give you a first customer and a way to prove your service works. You have to figure out whether anyone will keep buying it when the experiment ends.

You should look at what happens when the model concludes. If you assume that happens in three years, or five years, who still writes you a check, and for what? How much of your product would you have to rebuild to sell to a commercial plan or a self-insured employer? If those answers are unclear, work on them while the model is still paying.

Another place you should look is at concentration in the payment mechanism. In 2022, Signify Health announced its exit from BPCI Advanced, citing CMS’s retrospective calculations that lowered target prices and reduced the opportunity for savings. Revenue from dozens of providers can still depend on one payment formula. That is a risk worth understanding. Don’t assume that, just because you have a lot of customers, you’re diversified. 

On the policy side, Congress could make this work better by asking for three things.

First, ask for evidence that learning changes decisions. When a model ends or is revised, CMS should explain which findings it is acting on, what it will carry forward, and what the next test is meant to resolve. Stopping a bad idea early can be valuable. Repeating it without addressing what went wrong is much harder to defend.

Second, require a fuller account of results. Where spillovers are plausible, Congress should ensure CMS arranges access to other payers’ data early enough to study them properly. The agency should report clinical outcomes, access, and net costs alongside the federal spending result, and follow useful changes after the model ends. Section 1115A already allows CMS to require evaluation information from participants, though claims from other insurers will require additional arrangements. 

Third, Congress should ask that CMMI promulgate terms that people can reasonably plan around. Participants should understand how benchmarks will be calculated and revised before committing staff and capital. When CMS changes a model, it should explain why, give reasonable notice, and plan the transition for patients and providers. The ability to correct an experiment and the ability to build a business around it both matter.

The key to flight, as Orville and Wilbur discovered, is a combination of speed, lift and control. Today, CMMI is moving with a sense of urgency that deserves support. What we have been missing is the upward momentum and directional precision that can come from faster learning and better evidence. 

Congress has a chance to demand that. The next time there’s a hearing on the progress of CMMI, members of Congress should spend less time focusing on how many models have been judged a success under a narrow statutory definition. They should spend more time asking what the other experiments taught us, where their useful parts went, and how those lessons changed the models now being launched. We, the taxpayers, paid for that knowledge. We should insist that it’s put to use.