Misata vs Tonic Fabricate
Fabricate is the strongest of the new AI data generators: chat with an agent, get a fully relational database, export to more targets than Misata currently supports. The difference is what powers the rows. Fabricate generates through an AI model and meters usage per turn; Misata constructs rows with a deterministic engine, which is what makes three things possible that matter for test data: declared aggregates that land exactly, byte-identical regeneration from a seed, and unmetered generation. If you want an agent to improvise a database and you export to Oracle or Databricks, Fabricate is a serious tool. If your data needs to hit numbers you chose, reproduce exactly, and stay free to regenerate, that is Misata.
| Capability | Misata | Tonic Fabricate |
|---|---|---|
| Relational output with intact foreign keys | Yes, with per-relationship proof on every run | Yes |
| Declared outcomes hit exactly (curves, rates, totals) | Yes, constructed and verified per run | Not a documented guarantee; generation is model-driven |
| Reproducible output | Same seed, identical bytes, forever | Not a documented guarantee |
| Generation cost | Unmetered; the engine is deterministic math | Metered per generation turn on top of the subscription |
| Export targets | CSV, Parquet, SQL, DuckDB | Broader: Postgres, MySQL, Oracle, Databricks, Parquet, Avro |
| Design interface | Sentence, visual canvas, or your existing schema (dbt, Prisma, database) | Chat with an agent |
| Pricing | Free beta, proposed $5/month | Free with $10 monthly credits, then $29/month plus metered usage |
| Open source | MIT engine, runs fully local | No |
| Company behind it | Independent, solo-built, open | Tonic.ai, established and well funded |
A model that improvises versus an engine that constructs
Fabricate's generation runs through an AI model: you describe what you want, the agent produces it, and each generation turn is metered. That approach is flexible, and the relational integrity is real. What it does not advertise is exactness: ask for two million in December revenue and you get something near it, shaped by the model. Misata inverts the architecture. The declared outcome is the input to a deterministic solver, so the monthly totals land on the number to the cent, the same seed reproduces the same bytes, and regenerating costs nothing because no model runs in the data path.
What metering means for test data
Test data gets regenerated constantly: every schema tweak, every CI run, every demo reset. Per-turn billing turns that habit into a meter. Misata's generation is unmetered by design, because deterministic construction is essentially free to serve. The only thing Misata ever meters is the optional AI step that turns a plain-English sentence into a schema, and even that accepts your own key.
Where Fabricate is genuinely ahead
Fabricate exports directly to more databases, including Oracle and Databricks, and it can also work from existing data sources, which Misata's studio does not do in the browser. It is backed by an established company with an enterprise platform around it. If those are your constraints, it is the safer corporate choice today. Misata's bet is different: open source, provable, reproducible, and priced for individuals and small teams first.
Frequently asked
Both claim referential integrity. What is actually different?
The integrity claim is comparable; the difference is exactness and reproducibility. Misata additionally guarantees that declared aggregates land exactly and that a seed reproduces identical bytes, and it shows a per-relationship orphan count on every run so the integrity is verified rather than assumed.
Is Misata cheaper than Tonic Fabricate?
For generation-heavy use, yes. Fabricate is $29 per month plus metered per-turn usage after a $10-credit free tier. Misata's generation is unmetered, the beta is free, the proposed paid tier is $5 per month, and the engine is open source for unlimited local use.
When is Fabricate the better choice?
When you need direct export into Oracle, Databricks, or MySQL, want to derive data from existing sources, or your organisation prefers an established vendor. Those are real advantages today and pretending otherwise would be dishonest.