Gretel is gone. Here's how to choose what replaces it.

If you built on Gretel, you've noticed: gretel.ai now redirects to nvidia.com, the GitHub organization was archived in February 2026, the self-serve free tier is closed, and the pricing page doesn't exist. NVIDIA acquired Gretel in March 2025 for a reported $320 million-plus, and the technology now lives inside NVIDIA NeMo as two microservices — Data Designer (the LLM-driven synthetic dataset builder) and Safe Synthesizer (the privacy-preserving tabular synthesizer) — sold through NVIDIA AI Enterprise, sales-gated, with a migration guide for the standalone library.

That's a fine outcome for NVIDIA customers and a real problem for everyone else: a developer-first, credit-card-and-API product became an enterprise platform component overnight. If you're one of the teams deciding what to do next, this guide is for you. It is not a "we're the best" page — the right replacement depends on what you were actually using Gretel for, and for some of you the honest answer is "stay, and go through NVIDIA."

Facts about vendors below were verified on August 28, 2026. This market moves; check current documentation before you commit.

First: which Gretel were you using?

Gretel was really three products under one name, and each has a different natural successor.

1. Privacy-safe tabular synthesis — you pointed it at a table or a database extract and got back rows with the same statistical shape and no real individuals. This is the lineage that became Safe Synthesizer, and it's the crowded part of the market with the most alternatives.
2. LLM-designed datasets — you described the data you wanted (or seeded it with examples) and Gretel generated text, records, or instruction-tuning data with a language model. This became Data Designer, and it has fewer direct alternatives; most "synthetic data platforms" don't do it at all.
3. The API in a pipeline — Gretel as a step in a dev/test or ML workflow, called from code. Here the question is less "who has the best model" and more "who has a real API, SDK, and deployment model that fits your environment."

Sort yourself into one (or two) of those, and the shortlist gets short quickly.

The alternatives, honestly

NVIDIA NeMo Data Designer + Safe Synthesizer — the official path

Best for: teams already on NVIDIA AI Enterprise, or who used Gretel primarily for LLM-designed datasets — Data Designer is the direct continuation of that capability, and it's now open-source on GitHub with the managed service behind it.
Watch for: no self-serve, sales-led pricing, and a microservices deployment model. If you liked Gretel because a developer could start alone on a Tuesday afternoon, that's the part that's gone.

MOSTLY AI — privacy-safe tabular, no-code and SDK

Best for: privacy-first tabular synthesis for analytics and data sharing, especially in regulated European institutions, where it has a long track record. A no-code app for analysts, plus an open-source SDK for engineers.
Watch for: it's tabular-focused by design; if you need unstructured text generation, it's not the tool.

Tonic.ai — test data first, AI data second

Best for: engineering teams that need de-identified, referentially intact copies of production databases for dev and QA, with unstructured-text de-identification alongside. If "test data management" is the job, it's the most complete of this list.
Watch for: the center of gravity is test data; treat its ML/analytics synthesis as a secondary capability and evaluate it on your own utility metrics.

K2view — entity-based test data for large enterprises

Best for: organizations with data sprawled across many systems who need synthetic and masked test data that stays consistent across those systems (the same synthetic customer everywhere).
Watch for: it's an enterprise platform with an enterprise footprint and sales cycle — a lot of machinery if your need is one table.

Syntho — deep-learning "twins" with a European compliance focus

Best for: healthcare and financial-services teams in the EU that want a privacy-safe twin of a dataset with GDPR-shaped assurances and a hosted or on-premises option.
Watch for: smaller vendor; do your own utility and privacy measurement rather than relying on the twin framing.

YData — profiling plus synthesis for data teams

Best for: data-science teams that want data quality profiling and synthetic generation in one workflow (the ydata-profiling library is widely used), with a developer-oriented platform on top.
Watch for: developer-oriented is a plus or a minus depending on who's holding the tool.

SDV (Synthetic Data Vault) — open source, with a license caveat

Best for: prototyping, learning, and evaluation. The library is mature, well documented, and free to use for non-production work; SDMetrics, its evaluation library, remains MIT-licensed and is worth using no matter which platform you pick.
Watch for: most of the SDV ecosystem moved to the Business Source License in 2025, which restricts production use without a commercial agreement. It's no longer the "just pip install it and ship" option many teams assumed.

Syntheticore — measured privacy, hands-on onboarding

This is us, so weigh it accordingly. We generate privacy-safe tabular synthetic data — including multi-table datasets with keys and joins intact — through a web app and a REST API, deployable as managed SaaS or inside your own cloud. Two things we'd point at: every run ships a utility and privacy report (train-synthetic/test-real, distance-to-closest-record, membership inference), so "is it safe and is it useful" is a number, not a claim; and we pair the platform with people who help land the first dataset and hand the capability back to your team.
Watch for: we don't do LLM-designed text datasets — if that's the Gretel you used, look at Data Designer.

A scorecard, not a feature grid

Feature tables age badly and every vendor can tick every box in a demo. Score your shortlist 1–5 on the seven questions from our buyer's guide, on your data:

1. Does it handle your schema — multiple tables, keys, mixed types, skew?
2. Can it prove fidelity on your downstream task (train-synthetic, test-real)?
3. Can it prove privacy with numbers you can put in front of a reviewer?
4. Where does it run, and who sees the source data?
5. Does it fit your workflow — real API, SDK, pipeline step?
6. What's the total cost, including the engineering weeks to first trusted dataset?
7. Who helps when the first dataset is messy?

Add an eighth for this situation: what happens if this vendor is acquired too? Ask about data export, whether models you train are yours, and what a wind-down would look like. Gretel's customers didn't lose their data — but they did lose their pricing, their free tier, and their roadmap in a single announcement.

A migration checklist

Whichever way you go:

• Export everything first — configs, model artifacts, generated datasets, and the evaluation reports you used to justify the original decision. Archived repos and redirected domains have a way of losing documentation.
• Re-baseline on real data. Run your downstream model on the real training set and record the number. Every candidate gets judged against that, at the same row count — not against Gretel's old output.
• Re-measure privacy on the new tool, don't inherit it. Distance-to-closest-record, nearest-neighbor ratio, and a membership-inference test on the new generator's output (see how to measure synthetic data). Privacy is a property of the generator and the data, not a vendor attribute.
• Run two candidates in parallel on one painful dataset for two weeks. The scorecard above is filled in by doing, not by reading.
• Update your data-protection documentation. If a DPIA, a security review, or a customer contract named Gretel, it needs to name the successor and the new processing location.

The bottom line

Gretel's absorption into NVIDIA is the clearest sign yet that synthetic data has become infrastructure — important enough for the biggest company in AI to buy the category leader and fold it into the platform. For NVIDIA-aligned shops, the official path is reasonable. For everyone else, the good news is that the tabular-synthesis market is deep, the evaluation tooling is open, and the way to choose hasn't changed: measure utility and privacy on your own data, in your own environment, and pick the tool that gives you numbers rather than promises.