How OpenBill works out its numbers

What OpenBill shows

Search for a drug and OpenBill shows, for each product:

• What pharmacies pay. The national average price retail pharmacies pay to buy a drug from their suppliers (their acquisition cost), per unit, from a CMS survey of pharmacy invoices. The figure is CMS’s NADAC (National Average Drug Acquisition Cost).
• A Medicare price for the same drug. What Medicare Part D spent per unit in its latest data year (2024 in this build), or else the Medicare Part B payment limit for the current quarter, converted to the same unit. Which one is used is a rule below.
• The spread. The Medicare price divided by what pharmacies pay, stated as a multiple in one plain sentence.
• A cost-plus estimate. What a pharmacy would charge if it added a 15% markup and a $5.00 fee to what pharmacies pay. It is always estimated, and it is a model, not anyone’s price (the rule).
• What you paid, if you type it. Your own amount, as a multiple of what pharmacies pay and of the cost-plus estimate. It stays in your browser (the rule).

Every one of those numbers carries its state, the file it came from and its dates.

What it does not show:

• Your price. NADAC is what pharmacies pay their suppliers, not what a patient pays at the counter. Medicare’s figures are what Medicare paid or pays. What you pay depends on your pharmacy, your insurance, your deductible and the quantity you are given.
• A quote. A price here is a range to expect, read from government files that each have their own date.
• Cash prices, pharmacy list prices or insurers’ prices. No government file publishes them. The companies and pharmacies that do publish cash prices forbid storing or copying them, so OpenBill shows none. The cost-plus estimate is a calculation from NADAC; it is not a price any pharmacy charges.
• Medical advice. Nothing here says which drug to take or where to buy it.

There is no warranty. The figures come from other people’s data, which can be wrong, and from rules that can have mistakes. If you find a wrong number, tell us.

How to read a card

A card is one product: every NDC (the 11-digit National Drug Code printed on the package) that NADAC lists under the same description. This build has 5,936 products made of 30,031 NDCs.

Every number on a card has a state label and a provenance line, always visible: the source, the file it was read from, the date it applies from (the effective date) and the date OpenBill downloaded the file (retrieved). A figure that covers a period, such as a Part D data year or a Part B quarter, shows the period; its effective date is the first day of that period. A figure OpenBill calculates from others, such as a spread, shows each input’s own date or period beside it (for example “NADAC eff. 2026-09-23 · Part D 2024-01-01 to 2024-12-31”), never only the oldest. The provenance panel, opened from the card, adds the reason for the state, the NADAC explanation codes, the NDCs and the formula.

When a figure is missing the card says “We don’t know yet” and why. It never shows a zero instead. When a source is older than its limit (freshness), the card labels it stale and still shows it.

The four states

Strongest first:

filed

Read from a published government file (NADAC, ASP, a fee schedule). Strongest.

published

From a payer's or provider's own published rate.

derived

Computed from a published rule (e.g. Medicare pays 80% of the fee schedule).

estimated

Inferred, with the reason recorded.

In this build all 38,503 records in data/drugs.json are filed: each is a number printed in a government file. The numbers the page calculates from them, such as a Part B price per NADAC unit or a spread, are derived or estimated: of the 19,302 spreads on cards, 25 are derived and 19,277 estimated, for the reasons in the rules. A pack total is the exception: multiplying a price by your quantity is arithmetic, so the total keeps the state of the price it multiplies.

An example from this build

ARIPIPRAZOLE 10 MG TABLET, 31 NDCs:

What pharmacies pay

$0.11173 per EA filed CMS NADAC · nadac-national-average-drug-acquisition-cost-20261014.csv · effective 2026-09-23 · retrieved 2026-10-17

For 30 EA

$3.35 filed (the state of the price it multiplies: pack total)

Medicare Part D average spending

$1.54994 per EA filed Medicare Part D · DSD_PTD_RY26_P04_V10_DY24_BGM.csv · 2024-01-01 to 2024-12-31 · retrieved 2026-10-18

The spread

Medicare paid 13.9× what pharmacies pay. estimated (why it is estimated or derived)

Cost-plus estimate for 30 EA

$8.85 estimated (how it is worked out)

The rules

These are the rules in src/core/rules.js, in the words the page uses. Every number OpenBill calculates links to its rule here.

How prices are presented

OpenBill holds every screen to these five rules, word for word:

• A price is a range to expect, never a quote.
• Medicare's price is not your price. Say so on every screen that shows it.
• State which dataset and which effective date produced every figure.
• If a source is stale or a file is empty, show that rather than a zero.
• Disclaimer footer: not medical advice, no warranty, other people's data can be wrong.

The drug pages

On the website, each product also has a page of its own, and there is a page for each drug class and an A to Z index. These pages are built from data/drugs.json with the same rules as the card each time the site is published; they are not kept in the repository. They have no script. Their numbers are in the page itself, each with its state and source line, and each page links to the interactive card for its product, where you can change the quantity or enter what you paid. Start from the A to Z index of every drug.

• Drug classes are the U.S. Food and Drug Administration’s Established Pharmacologic Classes (the “[EPC]” entries of PHARM_CLASSES in the FDA NDC Directory), read for each NDC’s exact package. A product’s classes come from its NDCs that have an FDA identity. In this build 24,902 of the 27,083 NDCs with an FDA identity have classes listed, 24,086 of them an established class; there are 329 classes, and 4,797 of 5,936 products belong to at least one. Where FDA lists none, the page shows no class rather than a guess.
• Other strengths, forms and brands are the products that share an FDA nonproprietary name (the generic name on the label) with this one. They stay separate products. A drug page lists up to 24 of them, itself included, in a table with each one’s FDA dosage form, whether NADAC classes it brand or generic, what pharmacies pay per unit and its Medicare figure per unit, each with its state and source line. Products under one Part D name show the same Part D average, because Part D gives one figure for all of them.
• Compared with its class. For each of a product’s classes, its page says how many of the class’s other products priced in the same unit (EA, ML or GM) have a higher or lower NADAC price, and which product is in the middle. Each product is represented by the middle of its own NDCs’ NADAC prices, and the class by the middle of those; when there is an even number of prices, the lower of the two middle ones is taken, so the price shown is always one real filed NADAC price, never an average of two. A product whose NDCs are priced in different units, or whose class has fewer than 3 other products in its unit, gets no comparison. A unit is one tablet, capsule, milliliter or gram whatever its strength, so this compares prices per unit, not per dose.
• Shared explanations live here. What NADAC and Part D measure, how the starting quantity, the spread and the cost-plus estimate are worked out, and the rules every price follows are the same for every product, so a drug page links to them on this page instead of repeating them; the page keeps what is particular to its product. Its questions answer only what the rest of the page does not already say.
• Related drugs are up to five other drugs in the same classes, ranked by how many classes they share and then by Medicare Part D claims. A class says how a drug works. It does not mean one drug can replace another; that is a question for a doctor or pharmacist.
• Structured data. Each page describes its figures to search engines as costs of the drug (schema.org DrugCost), never as an offer to sell it, and every number there equals the one on the page.

Sources

Every number comes from one of these files. data/drugs.json records each file’s sha256, so you can check that you have the same bytes. The licences and terms, the fields read and the caveats of each file are in DATA_SOURCES.md.

How the files are joined

The NADAC file is the spine: one row per NDC in its newest week, 30,031 rows in this build. The other files are joined to it on exact keys. A join that is not certain is left empty rather than guessed; each empty join is counted below and explained on the card.

Identity: the FDA NDC Directory

Each NDC is looked up, digit for digit, in the FDA NDC Directory’s package list. FDA writes package NDCs with hyphens in three layouts (4-4-2, 5-3-2 and 5-4-1); the short segment is padded with a zero to make the 11 digits NADAC uses. A match gives the brand and generic names, dosage form, route, strength, labeler and package description.

27,083 of 30,031 rows (90.2%) match. The other 2,948 rows have no identity: their NDC is not in the directory’s package list (discontinued products leave it, for example), its package has no product record, or its FDA entries conflict with each other (0 NDCs in this build).

For the Part D name join only, a row with no identity can be matched using FDA’s record for its 9-digit product code, when that code has one product record and a sibling NDC of the same description is in the directory, or the NADAC description begins with FDA’s trade name or prints its name and strength. That record is not an identity: the row’s identity stays empty. Its route is used for one thing only: the multi-route rule below, for a Part D name marked *. In this build it was tried for 2,035 rows, and 1,385 of them joined a Part D name.

Medicare Part D: by name

Part D Spending by Drug has no NDCs. CMS lists each drug by brand name and generic name, with every strength and form together, so this join is by name: the NADAC description and, when the NDC has an FDA identity, FDA’s brand and generic names, against the names CMS uses. It is strict on purpose. A row gets no Part D figure when:

• the release forms differ: markers such as ER, XR or DR must be the same on both sides, and FDA’s dosage form must agree;
• more than one Part D row could match;
• the generic names disagree: the substances must be the same, and so must the salt where both names state one (a hydrate that is the only qualifier counts as the salt);
• a brand product has no Part D row with its own brand name;
• Part D lists the substance under several salts, or lists several presentations such as pens and vials, and neither the NADAC description nor FDA’s names say which one this NDC is;
• a combination product’s NADAC name lists a different number of ingredients from the Part D generic name;
• the product is sold over the counter under an FDA OTC monograph, or matches only a Part D row named by the generic name: Part D excludes over-the-counter drugs, so those Part D rows are a different, prescription product;
• NADAC prices the NDCs joined to one Part D name in different units and this NDC’s unit is not the one nearly all of them share, because CMS’s average would then mix tablets with milliliters;
• the Part D name carries CMS’s *, which marks one average over several routes of administration (tablets and injections, say), and this NDC is not on the route the name’s rows share. A row joins such a name only when its FDA route is known and includes the route that covers the most of the name’s rows, so every row that keeps the figure shares one route. The route is the one in the row’s FDA identity. A row the FDA NDC Directory does not list as a package of its own takes the route of FDA’s record for the same product (labeler and product code), and only of the record the name join already accepted for it with the checks above: one product record for the code, and either a listed sibling NDC of the same description, FDA’s trade name at the start of the description, or FDA’s name and strength printed in it. Such a row follows the route the rows with an FDA identity choose and does not count toward choosing it, because it is often a second package (a blister or pouch inside a carton) of a product whose listed package already counts; it keeps the figure only when its route includes the whole chosen route. Only when none of the name’s rows has an FDA identity with a route do those rows choose among themselves. A row with no known route, or on another route, is left out; if two routes cover as many rows but not the same rows, every row is left out. Routes are FDA’s own words, so PARENTERAL is not merged with INTRAVENOUS.

In this build 263 rows with no FDA identity took their route from the product record for a name marked *, and 146 of them kept the Part D figure because of it. Each of those rows says so in its provenance panel, naming the route, FDA’s product NDC as FDA writes it and, where FDA prints this NDC inside one of the product’s listed packages, that package; the name’s reason says how many of its NDCs took their route this way. The others were left out: 58 priced in a unit that is not the one nearly all of the name’s rows share, 55 on another route than the one the name’s rows share, 4 in a name whose rows tie between routes. 108 rows still have no known route and stay out.

19,277 of 30,031 rows (64.2%) match, pointing at 1,458 Part D names. For 17,111 rows CMS’s brand name is the generic name itself; for 2,166 rows it is a trade name. The exact rules are in tools/build_drugs.py, and the cases they were written for are in DATA_SOURCES.md.

CMS flags 61 Part D names as outliers: the average is substantially affected by unusual records. The 742 rows pointing at them keep the figure in the data but do not use it as the reference price. The flag is blank for 1 name (1 row); a blank flag counts as not flagged, and the provenance panel says so in the figure’s reason.

Medicare Part B: by exact NDC

The CMS NDC-HCPCS crosswalk lists which NDCs are billed to Part B under which HCPCS code, and how many billing units one package holds. An NDC is joined to it digit for digit; an NDC on crosswalk lines that disagree with each other is left out. The HCPCS code then gives the payment limit and the coinsurance percentage from the payment limit file.

To compare a payment limit with NADAC, the package has to be measured in NADAC’s unit. The crosswalk does not say which unit its package size is in, so OpenBill accepts it only when FDA’s package description for the same NDC gives the same total (milliliters, grams, or a count of tablets, vials and so on) and the crosswalk’s two billing-unit columns agree with each other. A kit that NADAC prices per EA is counted as whole kits. Where the two billing-unit columns disagree, one is used only when FDA’s strength × the package amount ÷ the HCPCS dosage gives exactly that number. Otherwise the package size is shown as the crosswalk prints it, a number with no unit, and the NDC gets no Part B price per NADAC unit.

1,427 of 30,031 rows (4.8%) are on the crosswalk, under 138 HCPCS codes; 1,338 of them (93.8%) have the package unit reconciled and can be converted to a price per NADAC unit: 1,289 because FDA’s package description gives the same total as the crosswalk; 32 because the NDC is a kit NADAC prices per EA, so the count is of whole kits; 17 because the crosswalk’s billing-unit columns disagree and FDA’s strength × amount ÷ the HCPCS dosage settles them exactly.

A payment limit is one figure per code. 38 of the 138 codes have a single labeler (the company named on the label) on the crosswalk, and 77 rows point at them. For the others the limit averages every labeler’s product billed under the code (when a Part B spread counts as derived). CMS spells some labeler names more than one way on the crosswalk: 6 codes list one company under names that differ only in capitals, punctuation or a suffix such as Inc or LLC. OpenBill counts each company once, and the provenance panel quotes every spelling; 2 of those codes have a single labeler only once the spellings are counted together.

What the cards end up showing

With these joins and the reference rule: 18,535 rows (61.7%) have a Part D reference price, 767 (2.6%) a Part B one and 10,729 (35.7%) none. By product: 3,075 Part D, 180 Part B, 2,681 none, of 5,936. 571 rows with a Part D reference also show a Part B figure in the provenance panel.

Every Part D spread is estimated (18,535). Of the Part B spreads, 25 are derived and 742 estimated. 387 estimated spreads at or beyond 20× or 0.2× carry the large-gap warning; on those cards the headline is “These two figures may not measure the same thing” instead of the spread’s sentence, with the multiple beside it. The full tables are in DATA-QUALITY.md.

NADAC explanation codes

NADAC marks each price with one or more explanation codes. Their meanings below are from the NADAC data dictionary. Codes 2 and 4 mean the price was carried forward from an earlier file rather than set by the latest survey; OpenBill marks those rows and shows every code in the provenance panel.

Known limitations

• Part D is a yearly average for a whole name. CMS averages every strength, form and maker sold under one brand and generic name, weighted by claims, for data year 2024 (2024-01-01 to 2024-12-31). It is gross spending as CMS defines it: ingredient cost, dispensing fees, sales tax and vaccine administration fees, whether the plan or the patient paid them, and before the rebates drug makers pay back, which CMS may not disclose. So it can sit far from the NADAC of one NDC, and prices may have moved since 2024-12-31. That is why every Part D spread is estimated. 2,154 rows point at names CMS marks with *, an average across routes of administration that can have different unit prices. Only NDCs on the route the name’s rows share keep such a figure (the multi-route rule), but the average itself still includes the other routes’ spending, so a tablet can be compared with a figure that is mostly injections or gels.
• The Part D unit is assumed. CMS reports spending per dosage unit but does not say which unit each drug uses. OpenBill takes it to be NADAC’s unit (EA, ML or GM) for the NDCs joined to that name. Where NADAC prices those NDCs in more than one unit, only the NDCs in the unit nearly all of them share get the figure; 2,739 rows got none for that reason. A spread far from 1× can still mean the units differ, which is why estimated spreads at or beyond 20× or 0.2× carry a warning.
• Part B pays only in some settings. Medicare Part B pays for a drug like this only in specific settings, such as a drug given in a clinic or doctor’s office or used with certain medical equipment; most prescriptions filled at a pharmacy are paid by Part D. The payment limit file’s own description of each code is in the provenance panel. A code shared by several labelers has one limit averaged across all their products, so a spread against it is estimated (100 of the 138 codes here). Billing units are whole numbers, so a price per unit can be slightly off for odd package sizes.
• Some NADAC prices are carried forward. 955 of 30,031 rows (3.2%) carry explanation code 2 or 4: CMS carried the price forward from an earlier file instead of setting it from the latest survey. They are shown with their codes, not dropped.
• Low recall, by design. 19,277 of 30,031 rows (64.2%) match a Part D name, and 10,729 (35.7%) have no Medicare price at all. A wrong Medicare price is worse than none, so an uncertain match is left empty and the card says “We don’t know yet”. Brand and generic versions of a drug are separate products, because OpenBill has no RxNorm identity yet. The 2,948 rows without an FDA identity have only the NADAC description and, for some, FDA’s record for their product code to match on, and they join a Part D name marked * only when that product record’s route is the one the name’s rows share.
• NADAC is a survey average. It is what retail community pharmacies paid on average in CMS’s survey. A given pharmacy may pay more or less, and NADAC says nothing about what a patient pays.
• The cost-plus estimate is a model. It applies one online pharmacy’s published markup and fee to NADAC, which is a survey average of what retail pharmacies pay, not what that pharmacy pays. Its real prices can be higher or lower, it charges shipping, and other pharmacies price in other ways. So the estimate says what a cost-plus price could look like, not what you will be charged (the rule).
• Dates are UTC calendar dates. The page counts every age against today’s UTC date, read from the viewer’s device clock, so the time zone never changes an age; near midnight UTC an age can differ by a day from the viewer’s local calendar. A device clock set behind can make a file look dated tomorrow; its age is then shown as 0 days. DATA-QUALITY.md gives each source’s age on the day it was retrieved.
• Older browsers cannot open it. The page carries its data compressed (deflate-raw, then base64) and unpacks it with the browser’s own DecompressionStream, which reads that format from Chrome 103, Edge 103, Firefox 113 and Safari 16.4. An older browser cannot unpack it; the page then says so in plain words and shows no figures rather than wrong ones.

Reproduce it

You need Python 3.10 or later (standard library only) and Node 20 or later. From a copy of the repository (on Windows the Python command is usually python rather than python3):

python3 tools/fetch_all.py
python3 tools/build_drugs.py
node build.js
node build.js --check
node --test

fetch_all.py downloads every source into tools/cache/ and records each file’s sha256; with --offline it only checks the files already there. build_drugs.py reads the cache, never the network, and writes data/drugs.json. build.js writes openbill.html, this page and docs/DATA-QUALITY.md. Nothing generated contains a build time, so the same files give the same bytes. build.js --check fails if a generated file is out of date, and node --test runs the unit, data, spot-check, quality and link tests.

Government sites replace their files. To reproduce these exact numbers, your 5 source files must have the sha256 values listed under sources. A newer file gives newer numbers; that is a data refresh, not a mismatch.

Report a wrong number

Every card has a “Wrong? Tell us” link. It opens the wrong-price issue form on GitHub with the card’s details filled in. You can also open the wrong-price form yourself.

The most useful report names the source file and the row in it that shows the right value. A receipt or a price from a website is useful context, but it does not show that a government file was read wrong. Please don’t include your own or anyone else’s health information.

How reports are checked and corrected: CONTRIBUTING.md and GOVERNANCE.md.