Comparison · Verified September 2026
Looking for a Gretel.ai alternative?
Gretel.ai was acquired by NVIDIA in March 2025. The standalone gretel.ai SaaS is no longer operating; the technology now lives inside NVIDIA NeMo. Here is what changed, and where SynthForge fits versus the NeMo-rebranded products.
TL;DR
Gretel.ai's standalone product is gone. gretel.ai now redirects to NVIDIA, and the gretelai GitHub org was archived on 2026-02-18. The capabilities split into two NeMo microservices: Data Designer (schema-driven generation, with optional LLM-generated columns) and Safe Synthesizer (differentially-private synthetic data trained on a real seed dataset). If you used Gretel because you had a real dataset and needed privacy-preserving copies, NVIDIA NeMo Safe Synthesizer is the successor. If you used Gretel for greenfield schema-driven test data, SynthForge is a closer match: web-based, the first 20,000 rows of every dataset free, no GPU, and no enterprise contract.
Recent context
TechCrunch reported the NVIDIA acquisition on 2025-03-19. As of September 2026: gretel.ai 301-redirects to nvidia.com (through www.gretel.ai, ending at NVIDIA's synthetic-data-for-agentic-ai page); the gretelai GitHub organization (including gretel-synthetics) is archived, with 12 of its 30 repos sharing the 2026-02-18 archive date; gretel.ai/pricing returns 404. The legacy Gretel free tier (15 credits/mo, ~100K records, $2/credit overage) is no longer the live pricing model. NVIDIA NeMo Microservices are licensed via NVIDIA AI Enterprise (sales-gated, GPU-shaped pricing).
When NVIDIA NeMo (the Gretel successor) is the right call
- • You have a real sensitive dataset and you need a privacy-preserving synthetic copy with differential-privacy guarantees. NeMo Safe Synthesizer is the productized version of the Gretel Tabular DP-SGD workflow.
- • You are already on NVIDIA AI Enterprise or running NIM-based infrastructure, so NeMo microservices fit your existing stack.
- • You need to integrate synthetic-data generation into a larger NeMo / agentic-AI pipeline (data designer + RAG + fine-tuning).
- • You need PII detection / classification driven by the Nemotron model family.
- • You can wait hours, not seconds, for output (DP-SGD training time is real) and you have a meaningful seed dataset to train on.
When SynthForge is the right call
- • You do not have a real dataset to train on. You are pre-launch, building a schema, or generating test data for a new feature. NeMo Safe Synthesizer cannot help you here; SynthForge can.
- • You want a web-based product you can start on without paying (the first 20,000 rows of every dataset are free), with no enterprise contract, no GPU, and no AI Enterprise license.
- • You need output in seconds, not hours: SynthForge is parametric / sampler-driven, not training-driven.
- • You need direct multi-dialect SQL output (Postgres, MySQL, SQLite, SQL Server, MariaDB, DuckDB, CockroachDB). NeMo emits CSV/JSON/Parquet via API; dialect-aware DDL is not the focus.
Feature comparison
Verified against primary sources in September 2026.
| Feature | SynthForge | Gretel.ai (now NVIDIA NeMo) |
|---|---|---|
| Product status (September 2026) | Live and hosted | Gretel SaaS shut down. Capabilities live in NVIDIA NeMo Microservices. |
| Requires a real seed dataset | No | Safe Synthesizer: yes (DP-SGD trains on it). Data Designer: optional. |
| Differential privacy | No | Yes. Safe Synthesizer with configurable epsilon, group-level / user-level privacy |
| Time to first output | Seconds | Data Designer: minutes. Safe Synthesizer: hours (training-bound). |
| Hosting model | Hosted SaaS. First 20,000 rows per dataset free, then credits. | NVIDIA AI Enterprise license (sales-gated, GPU-shaped pricing). gretel.ai SaaS is gone. |
| Free tier | Yes: the first 20,000 rows of every dataset | Legacy Gretel free tier (15 credits/mo) no longer applies. Current NeMo free-tier status: unverified; assume enterprise-gated. |
| Multi-table FK preservation | Yes (single-column, by construction) | Legacy Gretel Relational supported it. Under the NeMo rebrand: capability appears regressed or not yet exposed (unverified). |
| Output formats | CSV, SQL (7 dialects), JSON, JSONL, Parquet | CSV, JSON, Parquet via NeMo API. SQL dialect-aware output: not a focus. |
| Privacy report / metrics | Privacy benchmarks (DCR, NNDR) for evaluation | Synthetic Quality Score (SQS) and Data Privacy Score (DPS) on Safe Synthesizer output |
| PII detection | No | Yes. Nemotron-PII model lineage |
| Open source | Not OSS today | Legacy gretel-synthetics archived 2026-02-18. Active OSS is now github.com/NVIDIA-NeMo/DataDesigner. |
Pricing comparison
SynthForge
All features. The first 20,000 rows of every dataset are free, no credit card. 10,000,000 rows per table, and a 1 GB cap on the download.
500 credits for $5. About 12 datasets of 1,000,000 rows, or many more at typical sizes.
Gretel.ai (now NVIDIA NeMo)
The gretel.ai/pricing page no longer resolves. Free tier (15 credits/mo) and $2/credit overage are no longer in effect.
Sales-gated, GPU/subscription based via NVIDIA AI Enterprise. Public list pricing not surfaced. Treat as enterprise procurement.
What SynthForge does not do that Gretel.ai (now NVIDIA NeMo) does
Honest tradeoffs, in case they decide the comparison for you.
- • If you need differential privacy, SynthForge does not have it. NeMo Safe Synthesizer does. This is the cleanest case for picking NeMo over SynthForge.
- • If you have a real sensitive dataset and you need to train a model on it for a privacy-preserving synthetic copy, SynthForge is not designed for that workflow.
- • If you are building agentic-AI pipelines on NVIDIA NIM, NeMo's ecosystem fit is a real advantage SynthForge cannot match.