Best Data Labeling Tools 2026: Label Studio, CVAT, Labelbox
The short answer
Label Studio for most teams, CVAT for computer vision, Labelbox when you want a managed enterprise pipeline. As of November 16, 2026:
- Label Studio — best overall. Apache-2.0, multimodal, strong for LLM evaluation; self-host free or pay $99/month for hosting.
- CVAT — best for images, video and 3D. MIT licence, SAM 2/SAM 3 segmentation, cheap hosted plans.
- Labelbox — best managed platform. Usage-based Labelbox Units (LBUs), 500 free a month.
- Prodigy — best for a single NLP developer. Self-hosted, lifetime licence from $390.
- Argilla — free option for text and LLM feedback, with slower releases.
The comparison
| Tool | Licence | Hosted price (Nov 16, 2026) | Data types | Best for |
|---|---|---|---|---|
| Label Studio (HumanSignal) | Apache-2.0 Community Edition; 1.23.2 (Sep 29, 2026) | Starter Cloud $99/month, +$49 per extra user, up to 12 users; Enterprise custom | Text, images, audio, video, time series, LLM responses | Mixed workloads, LLM evaluation |
| CVAT | MIT; v2.83.0 (Nov 12, 2026) | Free; Solo $33/month ($23 yearly); Team $33 per user/month (2-seat minimum, $46 per 2 seats yearly); Enterprise self-hosted | Images, video, 3D point clouds, audio | Computer vision |
| Labelbox | Proprietary SaaS | Free 500 LBUs/month; paid plans billed per LBU; Enterprise contract | Images, text, chat, audio, PDF, video | Managed enterprise data factories |
| Prodigy (Explosion) | Commercial, self-hosted | $390 personal lifetime licence; $490 per company seat in packs of 5 | Text (NER, classification), images, audio, video | Solo NLP developers, active learning |
| Argilla | Apache-2.0; PyPI 2.8.0 (Mar 2025) | Self-host free | Text, LLM feedback | Small text and RLHF datasets |
1. Label Studio — best overall
Label Studio’s strength is that the labeling interface is a configuration, not a fixed product: the same tool can draw boxes on images, tag spans in text, transcribe audio or show a prompt with two model answers and ask a reviewer which is better. That makes it the default for teams whose work moved from classic annotation to LLM evaluation and preference data in 2025–2026. The Community Edition is Apache-2.0 and actively maintained (28,700 GitHub stars; commits on November 16, 2026). HumanSignal’s hosted Starter Cloud costs $99 a month plus $49 per additional user, up to 12 users, with role-based access, quality workflows and automated task distribution. Enterprise adds SSO, SOC 2 and HIPAA reporting, LLM-as-a-judge, auto-labeling and bulk labeling.
2. CVAT — best for computer vision
CVAT is the specialist for pixels: bounding boxes, polygons, tracks across video frames and 3D point clouds, with SAM 2 and SAM 3 segmentation for interactive masks. It is MIT-licensed and ships often (v2.83.0 on November 12, 2026). CVAT Online is unusually cheap: Free includes the core editor, ground-truth and honeypot QA, API access and export; Solo is $33 a month or $23 billed yearly and adds batch auto-annotation and full dataset export; Team is $33 per user per month, sold in two-seat packs, with shared workspaces, SSO and audit logs. Without a Team plan, an organisation workspace is evaluation-only (2 members, 1 project, 3 tasks).
3. Labelbox — best managed platform
Labelbox bills in Labelbox Units (LBUs), a normalised credit consumed differently by its Catalog, Annotate and Model products. Annotating costs 1 LBU per data row for images, text, chat or audio, or 1 LBU per page for PDFs; storing data in Catalog costs 1 LBU per 60 data rows a month. Free accounts get 500 LBUs a month; when they run out you can still export, but cannot add rows or labels until the next period. Choose it when you want curation, labeling, model-assisted pre-labels and outsourced labelers in one contract rather than a tool you operate.
4. Prodigy — best for one NLP developer
Prodigy is a self-hosted Python tool from the makers of spaCy, built around fast keyboard-driven annotation and active learning. It is a one-time purchase: $390 for a personal lifetime licence with 12 months of upgrades, or $490 per seat for companies in packs of five. Data never leaves your machine.
5. Argilla — free text and feedback datasets
Argilla (Apache-2.0, now part of Hugging Face) is a simple way to collect human feedback on text and LLM outputs and push datasets to the Hugging Face Hub. The repository is still active, but the last PyPI release is 2.8.0 from March 2025 — fine for a small project, a risk as a long-term platform.
How to choose
- LLM evaluation, preference data, mixed media: Label Studio.
- Images, video, 3D: CVAT.
- Large programme with outsourced labelers and procurement: Labelbox.
- One developer labeling text locally: Prodigy.
- Before buying any of them: decide what you can generate synthetically — see best synthetic data generation tools — and how labels feed your evals: how to run LLM evals in CI/CD.
Last verified: November 16, 2026.