Search for gretel.ai today and you get redirected to nvidia.com. NVIDIA acquired Gretel in March 2025 for reportedly north of its $320 million valuation and relaunched it as NeMo Data Designer under Apache 2.0. That single redirect captures the last eighteen months in this space pretty well: the venture-backed vendors got bought or repriced, and the open-source tooling got noticeably better.
The change that actually affects a decision you're making this week is quieter. SDV, the most widely cited open-source synthesizer, left the MIT license behind for the Business Source License. I wrote up the full comparison, with every license and price checked against the vendor's own page in August 2026, on DevToolLab - this is the condensed version, plus the local test I'd run before trusting any of these tools.
Three different jobs share one name
"Synthetic data" gets used for three distinct problems, and most of the bad tool picks in this space come from solving for one while actually needing another.
Mock data fabricates rows with no connection to anything real - names, emails, cities. Good for seed scripts and test fixtures. Statistical synthesis trains on your actual table and generates new rows that preserve the distributions and correlations between columns, which matters the moment something downstream learns from the output. De-identification takes real production data and scrubs it so individuals can't be re-identified while keeping row-level structure - a compliance problem, not a data-generation one.
The gap between the first two is the expensive one, and it's what the test near the end of this post is built to expose.
The open-source landscape, quickly
Faker is MIT, ubiquitous, and exactly as good as its job description: independent, plausible-looking fields. Seed it and you get byte-identical output across runs, which is genuinely useful for CI fixtures:
from faker import Faker
fake = Faker("en_US")
Faker.seed(42)
for _ in range(3):
print(f"{fake.name()},{fake.email()},{fake.city()}")
One gotcha worth knowing: the seed locks the entire random stream, not each field separately. Add one fake.* call anywhere earlier in the script and every value after it shifts. Pin the script alongside the seed.
SDV (Synthetic Data Vault) is still the deepest open synthesizer - relational tables with foreign keys, sequential data, built-in quality evaluation - but it's Business Source License now, not MIT, the same license family Terraform and Vault moved to. My infrastructure as code piece ran into the identical question. BUSL allows plenty, but check the terms yourself rather than trusting a roundup that still calls it "MIT licensed" - plenty do.
MOSTLY AI's SDK is Apache 2.0 and local-first by design: training and generation run on your own machine by default, no cloud account required. If SDV's license is a blocker, this is the first thing to try.
NVIDIA NeMo Data Designer - what Gretel became - targets generating training data for models rather than cloning a production table. Dependency-aware fields, Python/SQL validators, LLM-as-judge scoring on the output, and a preview mode before you generate at scale. Apache 2.0.
The commercial tier, in one sentence each
Tonic.ai splits into three products worth telling apart: Fabricate (mock generation, the only one with public self-serve pricing - free tier, then $29/month), Structural (de-identification against real databases, quote-based), and Textual (unstructured redaction, billed per word count). The pattern holds across the category - K2view, Syntho, YData are all quote-based enterprise deals. If you need a signed DPA and a live database connector, you're in procurement, not a credit card checkout.
The test that actually separates these tools
Averages lie here. Generate a real dataset with a deliberate relationship - salary rising with years of experience - then compare it against columns generated independently versus columns run through an actual synthesizer (SDV's GaussianCopulaSynthesizer, in my test):
real data corr=+0.988 mean_salary=$119,977 mean_years=20.3
independent columns corr=-0.061 mean_salary=$123,101 mean_years=20.4
SDV synthetic corr=+0.880 mean_salary=$118,319 mean_years=21.5
Every mean is within a few percent across all three. Any "does this look reasonable" check passes all of them. But the correlation between experience and pay goes from +0.988 to essentially zero once the columns are generated independently - the relationship isn't weakened, it's erased. A model trained on that data learns that experience has no bearing on pay. SDV holds the correlation at +0.880, and crucially, zero of its generated rows matched a real row in the source table.
Run this on your own data, with your own columns, before picking a tool based on a vendor's demo.
Picking one
Dev database or test fixtures: Faker, seeded, nothing else needed. Training or evaluating on data you can't touch directly: MOSTLY AI if you need a permissive license, SDV if you want the deeper feature set and BUSL works for you. Generating eval or instruction datasets for LLMs: NeMo Data Designer with preview mode and judge scoring turned on. Real production database plus compliance sign-off: a commercial platform on a quote, budgeted as procurement.
For quick one-off rows without installing anything, DevToolLab's Fake Data Generator runs in the browser, and the SQL Insert Generator turns the output into statements you can paste into a seed script.
Top comments (0)