How SpiceGrinder Compares

This page is here so you can work out which kind of tool you need before comparing the products inside it — and then the detail pages tell you honestly where we fit and where we don’t. After all, there are good tools in this space, and several of them are better than us at the specific thing they were built for, so we’ll be honest about that and state it where it matters.

Four approaches, not one market

“Synthetic data” covers four very different things. Most comparison articles mix them together, which often makes them difficult and unsatisfying to read.

Value libraries — Faker, Bogus, Mockaroo. These generate plausible individual values: names, addresses, emails, phone numbers. Fast, free or cheap, almost no setup. Each field is independent of every other field.

Model-driven generation — GenRocket, DATPROF, SpiceGrinder. You describe the structure, distributions, and relationships; the tool generates data matching that description. No real data required as input. This is our category.

Generative / ML platforms — Tonic.ai, Gretel (acquired by NVIDIA in 2025), MOSTLY AI (ceased operations in 2026). Train on a real dataset, then sample new rows that resemble it statistically. Powerful when you have a reference dataset and the access to use it.

Domain simulators — Synthea. Simulate a specific domain in depth using expert-built models. Synthea’s clinical modules are built by clinicians and calibrated against CDC and NIH statistics, and it’s free.

So which category should you look for?

There really isn’t a single answer to “which tool is best?” The real questions to ask are what do I have? and what do I need?

If you pick a tool that’s not suited for the job, it’s not going to give you what you really need — regardless of how good the tool is. Most bad results in synthetic data tooling are the result of using the wrong tool, not any flaw in the actual products.

The detail pages

Found your category above? These are the head-to-head comparisons, each one honest about where the other tool wins.

vs. Faker, Bogus, and Mockaroo → Faker is better than us at producing a realistic phone number. But it has no concept of that phone number belonging to the same person who placed the order in the next table.

vs. GenRocket → The closest thing to a direct competitor, and a mature one. Starts at $55,000/year with no free tier (their published price, checked September 2026). Its modeling step runs in their cloud even though generation runs on your premises.

vs. AI-generated synthetic data → Training on production data means having production data. If that access (or existence) is your blocker, this approach relocates the problem rather than removing it. The output is a sample, so it’s different every run.

vs. Synthea → If you need realistic FHIR patient records, use Synthea — we can’t match MITRE’s clinical depth. The limit is that claim volume is tied to population size, and the distributions are Synthea’s rather than yours.

Where SpiceGrinder fits

You describe your data as a plain-text model — entities, distributions, relationships — and generate from it locally. Nothing is trained, nothing is uploaded, and nothing about your model leaves your machines, including while you design it.

Three properties follow from that, and all three are checkable rather than asserted:

• Deterministic. Same model, same seed, byte-identical output — on your laptop, in CI, on a colleague’s machine. Run it twice and diff it.
• No real data required. There’s no training step, so no copy of production has to exist for this to work. Not forbidden — if you have a sample, you can fit a model to it — but that’s a one-time analysis, never a standing dependency.
• Models are text. They live in your repo, diff in a pull request, and get reviewed like code.

Free is genuinely free — a separate build with no time limit and no row cap, though models are limited to 20 nodes and numeric output. Pro is $199/month, published, no quote required.

The longer version of that argument lives on Why SpiceGrinder.

The category is a starting point, not a fence

There’s one asymmetry in all of this: the other three categories describe products. Faker is a value library. Synthea simulates one domain. A generative platform is a trained-model service. SpiceGrinder is an engine — which is why it sits in the model-driven box by default though it is not confined to it.

Fitting an existing dataset is covered above: we ship an agent skill, fit-model-to-sample, that does explicitly and reviewably similar to what a generative platform does opaquely.

If no value library covers your data, build the library. Custom generators and filters are a documented extension point, and FlatFile turns any vocabulary you assemble into a reusable pool. Nobody ships a component for your firm’s instrument identifiers or your regulator’s claim codes — not Faker, not us — until someone builds it. Whatever you build and whatever it generates is yours; there’s no license on the output.

If your domain has no equivalent to Synthea, you can accumulate one. Import composes models from smaller reusable sub-models, so the fifteenth model you write starts from the library the previous fourteen helped create. The shipped samples already work this way: claims-core-lib.xml is a reusable frequency-and-severity library that claims-imported.xml pulls in, not a one-off script.

We’re not claiming to beat a specialist at the thing it specializes in. In your domain we might; we might not. The point is the intersection — one engine you can point at whichever of these your problem actually turns out to need, aimed at getting you the data you need, when you need it.

Consider combined approaches

The category map isn’t a set of walls. Several of these pair well with SpiceGrinder, and we built two of them end to end to show how:

Claims at volume, on a Synthea population → Synthea supplies the member spine; SpiceGrinder generates claims against it. One million claim lines in 1.6 seconds, versus roughly three hours out of Synthea alone — because volume and population become independent.

Multi-market order data, on Faker-built pools → Our bundled data is US-centric. Faker has dozens of locales. Faker supplies the vocabulary, SpiceGrinder supplies the market mix and the relationships.

If we’re not the right fit

We’d rather say so now than after you’ve spent a week on it. SpiceGrinder is the wrong tool if de-identifying an existing production database is the actual task, if you need clinical realism you’d rather not build yourself, if your data’s value lies in high-dimensional structure nobody has articulated, or if one flat table of plausible values is genuinely all you’re after.

Note what isn’t on that list: having a real dataset. Real data is never required by SpiceGrinder, but it’s perfectly usable — our fit-model-to-sample agent skill fits a sample statistically and hands you a readable best-fit model. The sample is read once, during that analysis. From then on, generation never touches it again.

If you need related, structurally consistent data generated from a description — reproducibly, locally, without sourcing anything sensitive to begin with — that’s what this was built for.

Download SpiceGrinder Free → No account, no email, no row limit.