2026 ACA Risk Adjustment Model: What Changed

Four things changed. One is being read as a bigger coverage change than it is, and one is an addition that will lower some members’ scores.

The 2026 package landed with little fanfare: a tables workbook, an instructions PDF, and for the first time Python software in place of SAS. Read the CMS summary and you get four bullet points. Read the tables and you get something more useful, and in two places something different.

We compared every cell of the CY2025 and CY2026 packages: all 404 model coefficients across five metal tiers, the full ICD-10 crosswalk, both drug crosswalks, the procedure code list, and the cost-sharing table. Here is what is actually in there.

1. A new class of variable

For the first time in years, CMS added a structural element rather than adjusting the existing ones. Affiliated Cost Factors capture significant costs not already associated with a condition category (HCC) or a prescription drug category (RXC).

Two exist for 2026, both for HIV pre-exposure prophylaxis. ACF_PrEP in the adult model fires when a member has a qualifying drug or procedure code and does not already have the anti-HIV drug category, worth 3.490 to 3.740 depending on metal tier. ACF_PrEP_Child fires from age 12, when a member has a qualifying code and does not already have the HIV/AIDS condition category, worth 1.914 to 2.143.

To put 3.609 (adult, Silver) in perspective: that is more than most single condition categories are worth. It arrives with two brand-new crosswalk tables, and (the part that matters operationally) it is identified from pharmacy and procedure data, not from diagnoses.

2. Every coefficient that matters moved

The model was recalibrated on 2020 to 2022 EDGE data, replacing the 2019 to 2021 blend. 320 of the 404 existing factors changed. The 84 that did not are all structurally zero, categories absorbed into group variables that never carried a coefficient in the first place. Essentially every factor that contributes to a score moved.

The direction is not uniform, and this is worth internalizing before anyone forecasts anything.

Adult books should expect downward pressure from recalibration alone. Child-heavy books should not assume the same.

3. The 77-code question

This is where the widely-repeated summary needs a correction.

CMS tightened the definition of an acceptable service for risk adjustment, and 77 codes came off the included list. That number is being quoted as though 77 services can no longer evidence a diagnosis. Our read of the data says otherwise.

Sort those 77 by clinical family and compare them against the 75 codes added in the same release:

• 49 look like genuine exclusions. Telehealth and online digital E/M (12 codes), remote monitoring and imaging (9), therapy and training (9), care management and preventive visits (6), hearing aid services (6), respiratory supervision (3), plus a handful of others. These share a characteristic: no face-to-face diagnostic encounter.
• 24 look like ordinary CPT churn. The lower-extremity revascularization family was renumbered: 10 codes out, 23 replacements in. Eight Category III neurostimulator codes converted to Category I. Two urology codes, two cardiac codes, two aortic codes: same pattern.
• 4 are genuinely ambiguous, and we have flagged them for review rather than guessing.

Treating all 77 as coverage losses overstates the impact by roughly a third, and sends clinical teams chasing a documentation problem that does not exist.

The distinction matters because the two groups need different responses. A genuine exclusion is a documentation and capture problem. A renumbered code is a mapping-table problem.

4. The decrease hiding inside the addition

Here is the finding we did not expect.

Across both drug crosswalks, no NDC moved between prescription drug categories. Every removal is a product leaving the crosswalk, every addition is a product new to it: 224 in, 294 out, pure RxNorm product churn. The drug class definitions are unchanged.

With one exception. Two HCPCS codes, J0738 and J0752, left the Anti-HIV Agents category and now appear in the new Affiliated Cost Factor crosswalk instead. CMS reclassified them from HIV treatment to HIV prophylaxis.

For most members this is invisible. But consider an adult whose only HIV-related evidence was one of those two codes.

That first member’s score falls by roughly three quarters of a point, not because of a data gap, not because of a coding failure, but because CMS moved the code. If you have not anticipated it, it will look like a defect in your implementation during the impact assessment. It is not. The second row is the population the factor was created for.

5. The operational chain

The changes above sit at different points in the pipeline, and they fail in different ways.

The acceptable-service list decides whether a claim line may act as a source of diagnosis. That feeds diagnosis capture. Capture feeds the condition categories. Categories feed the score. The score feeds gap targeting, provider scorecards and revenue forecasts.

So a change at the top propagates all the way down. And, the important part, nothing errors. A member whose only evidence for a chronic condition came through an excluded telehealth visit does not generate a warning. They simply score lower.

Re-running the score without re-running claim risk-eligibility for the full benefit year produces a number that looks correct and is not.

What plans should be doing now

• Check your pharmacy and procedure feeds are complete. They now drive three things: the drug categories, the drug-by-condition interactions, and whether the new cost factor fires at all. A gap does not error; it just scores low.
• Re-run claim risk-eligibility across the whole benefit year, not just new claims. Anything processed under the CY2025 list carries a stale flag.
• Identify your PrEP population before the impact run, so the reclassification reads as expected behavior rather than a bug.
• Version your reference data by benefit year. CMS has said the procedure, diagnosis and drug tables will all be reissued later this year. Anything hard-coded becomes a code deployment at the worst possible time.
• Separate the drivers when you measure the impact. Recalibration, the new factor, the service-code filter and the cost-sharing change all move the score at once. If you cannot attribute a movement to one of them, you have not finished.

A note on measurement

Every figure above is a model coefficient or a code count. None of it is a forecast of any particular plan’s risk score, and it should not be read as one: your movement depends on your membership and condition mix, and the only way to know it is to measure it.

The way we do that is a dual-run: the identical membership scored under both the complete CY2025 and CY2026 configurations, so nothing but the model differs between them. The total movement is then decomposed driver by driver, and every member above the materiality threshold is explained down to variable level. Movement we cannot account for blocks publication rather than getting rounded away.

One last thing worth flagging: the July package is not the final word. CMS has stated the procedure code list will be updated through October 1, 2026, the diagnosis crosswalk extended to FY2027, and the drug crosswalks refreshed again. Reference data continues to move through CMS’s technical channels between formal publications. Build for a data release, not a code release, because there will be another one.

Invent Health builds and validates risk adjustment models for ACA health plans. Every figure in this analysis was re-derived independently from the CMS source files and cross-checked between them before publication.