How much a marketing mix model costs depends mostly on how much human time comes with it: the software can be free, while the people, data and experiments around it run from a few thousand dollars to well over $200,000 a year. Among vendors that published prices in October 2026, Stella ran from free to $6,000 a month, Sellforte listed $4,500 a month for its full-funnel MMM at $300,000 of monthly media spend, and Paramark listed $100,000 to $220,000 or more a year. Enterprise consultancies don’t publish prices at all, and building in-house with open source swaps a license fee for a share of a data scientist, whose US median wage was $120,230 in May 2025 according to the Bureau of Labor Statistics.
I’ve been on both sides of this decision. I built marketing mix models for clients like Monzo Bank, Booking.com and The AA, and I’ve advised Google, Appsflyer and Recast on MMM, so I’ve seen what it takes to build a model from scratch and what vendors charge to save you the trouble. (I’ve advised Recast, which is mentioned once below, so take that as my disclosure.) The mistake I see most often is comparing a vendor’s annual fee against a free open-source library, as if the library builds itself. I’ll break down the real cost components first, then give you a calculator to compare the three options with your own numbers.
How much a marketing mix model costs, line by line
A marketing mix model has seven cost lines, and every option pays all of them in some form. What changes between building and buying is who pays which line, and whether it shows up on an invoice or in somebody’s calendar:
- Data engineering: getting weekly (or daily) spend, impressions, sales, prices, promotions and other controls out of a dozen systems into one clean table, and keeping it flowing. On most first projects this is the biggest single piece of work, which is why the MMM data requirements page exists.
- Analyst or data scientist time: specifying the model, choosing priors and controls, checking diagnostics, explaining results and arguing with stakeholders about them.
- Compute: running a Bayesian model takes real processing time. Google recommends at least one GPU for Meridian and says it’s tested on a T4 GPU with 16 GB of RAM, which is a modest cloud bill compared with salaries.
- Refresh cadence: every refresh means new data, a refit, a review and a conversation. Monthly refreshes mean three times as many refits and reviews as quarterly ones.
- Experiments: the spend you withhold or add in test regions, plus any test software, to calibrate the model. This cost is the same whichever option you pick, and skipping it is the most expensive saving you can make.
- Vendor fees: a SaaS license, usually annual and often scaled by media spend.
- Consultancy fees: usually an initial project to build the first model, then a fee per refresh or a retainer.
The first two lines are the ones people underestimate. When I’ve built models, the regression itself was a small fraction of the time; most of it went into the data and into turning a coefficient into a recommendation someone would act on.
What vendors publish about MMM pricing
Most vendors quote prices only after a call, but a handful publish them, and these are the real numbers available on 24 October 2026. Treat them as reference points rather than what you’ll pay, because most scale with media spend and channels:
| Vendor | Published price | What it includes |
|---|---|---|
| Stella | Free; $750, $3,000 and $6,000 a month | Self-serve; MMM and regional holdouts on connected data from $3,000. |
| Cassandra | From 2,700, 3,400 and 4,050 a month | Self-serve to managed; rises with media spend; currency not shown. |
| Sellforte | $4,500 a month for Full-Funnel MMM | At $300k monthly media spend; $8,500 for Enterprise. |
| Paramark | $100k, $150k and $220k+ a year | One to three or more MMMs, unlimited tests, a dedicated advisor. |
| GeoLift by Recast | $100 a month or $1,000 a year | Geo testing only, after a six-month free trial. |
| Northbeam | $1,500 and $3,500 a month | Attribution tiers; MMM+ is an optional add-on on Enterprise (custom price). |
| Funnel | From $300 or $600 a month | Data hub only; the Measure (MMM) add-on isn’t priced publicly. |
| Meridian, Robyn, PyMC-Marketing | Free | Open-source software; you supply the people and compute. |
Annualised, the self-serve SaaS options run from about $9,000 to $72,000 a year, the fuller MMM platforms from about $54,000 (Sellforte) to over $220,000 (Paramark Enterprise), and the open-source libraries cost nothing to license. The MMM software directory has the full list of 30 vendors, including the many that don’t publish prices, and the open-source MMM libraries page compares Meridian, Robyn and PyMC-Marketing.
MMM cost calculator: build vs buy over three years
The calculator puts all seven cost lines side by side for the three options. Only the salary default is a real figure (the BLS median for data scientists); every other default is a placeholder chosen to make the chart readable, so replace them with your own salaries and real quotes before you draw any conclusion. The input that moves the answer most is usually the FTE share, because people dominate the in-house cost.
Total cost over the years you chose, by cost line.
With the placeholder numbers, SaaS comes out cheapest and the consultancy most expensive, but small changes flip the order. If your data scientist only needs 15 percent of their time once the model is built, in-house edges ahead; if your vendor quote is at the Paramark end rather than the Stella end, it loses. That’s the honest answer to “how much does a marketing mix model cost”: there’s no market rate, only a set of trade-offs you can price.
What building in-house with open source really involves
Building in-house means choosing a library (Google’s Meridian, Meta’s Robyn or PyMC-Marketing are the main options), assembling the data, building and validating a first model, and then refreshing it on a schedule. With an experienced modeler and clean data, a first credible model is a matter of weeks; with messy data and a team learning the method, it’s months. After that the ongoing work is lighter, but someone still has to own each refresh, investigate when results move, and keep the data pipeline alive when an ad platform changes its API.
The advantages are control and transparency. You can see every assumption, change the model when the business changes, and you’re never locked into a vendor’s roadmap or contract term. The disadvantages are key-person risk (if your one modeler leaves, the model often leaves with them) and the fact that nobody outside your team has checked your work. If you go this route, write the model specification down and run at least one experiment a quarter, so the model meets ground truth regularly. The Python and Excel walkthroughs are a cheap way to find out how comfortable your team is before you commit.
When a SaaS platform is the better buy
A SaaS platform is usually the better buy when you don’t have a data scientist with spare time, when you want a model refreshed weekly or monthly without anyone babysitting it, and when your channel mix is fairly standard (paid social, search, some video, maybe TV). You’re paying for a data pipeline that someone else maintains, a model that someone else has already run for other brands, and an interface your marketing team can use without asking an analyst. The internal time doesn’t disappear, though: somebody on your side still needs to understand the outputs well enough to challenge them, and the vendor evaluation questions are the place to start.
Watch for two things in the quote. First, how the fee scales with media spend, because a license that’s cheap at today’s budget can double when you grow. Second, what you keep if you leave: your own data, the model’s outputs and ideally the history of its estimates, so a future vendor or in-house team doesn’t start from zero.
When a consultancy earns its fee
A consultancy earns its fee when the questions go beyond digital channel allocation: retail sales through distributors, price and promotion effects, brand investment, multiple countries, or a board that wants an independent view. The big firms bring benchmarks from many clients and experienced modelers who’ve seen your category before, and a Google-certified Meridian partner can give you a model built on open-source code that you could later take in-house. None of the large MMM consultancies publish prices, so the fees in the calculator are placeholders you’ll need to replace with a real quote.
When you get that quote, ask for the initial build and each refresh priced separately, and find out how much of the fee is data work you could do yourself. The cheapest consultancy engagement I can imagine is one where you arrive with clean data and a clear measurement plan, because you’re paying for judgment rather than spreadsheet wrangling.
How to decide between building and buying an MMM
I’d use the questions below as a quick screen before you price anything. If you tick most of the first four, building in-house is realistic; if you tick most of the last four, buy, and the remaining question is whether you need software or people.
Whichever way you go, budget for experiments as a separate line, because a model you can’t check is a model you’ll eventually stop believing. The incrementality testing tools page covers the cheapest ways to run them. And if you want your team to understand the method well enough to build a model or to hold a vendor to account, the Vexpower MMM course teaches it on an invented brand, Northstar, starting from what marketing mix modeling is.
