Labelbox Review

2.6 million contributors. 50K+ PhDs. Those aren't vanity numbers — that's the Alignerr Expert Network, and it's what separates Labelbox from Scale AI or CVAT in a head-to-head. When your model needs reward signals and preference pairs from credentialed domain experts, you can't fake that supply chain.

The product has clearly shifted upstream. Reinforcement Learning Data, Private AGI Benchmarks, Arena Evals — this isn't a labeling tool anymore. It's an AI training infrastructure play. That's the right bet right now. The tradeoff: if you just need fast, cheap image annotation for a computer vision team, the marketing site no longer speaks to you — the self-serve Starter plan at $0.10 per LBU is still there, but you'll find it in the docs rather than on a pricing page.

$189M raised, a business the company says is profitable, and research presence at CVPR and NeurIPS all point to a company with staying power. Pilot the free tier up to 30 users before entering a contract conversation.

The Alignerr Expert Network — 2.6M+ contributors, 50K+ PhDs, 85K+ licensed professionals across 200+ domains — is the kind of labor supply that takes years to assemble. Scale AI has comparable depth, but Labelbox's pivot toward knowledge work rubrics and RLHF preference pairs shows they're tracking where model training is actually headed, not where it was in 2021. That's a meaningful architectural bet.

The free tier caps at 30 users and 50 projects, which is generous for evaluation, and Starter lifts the user cap at a fixed $0.10 per LBU without a sales call. Enterprise pricing is contact-only, and so is the frontier work — Horizon, Terra and Recursion publish no price — which means budget visibility for the strategic bets arrives only in a sales cycle — that's the operational friction point every Head of Data needs to plan around.

If we adopt this, in 3 years we've either built a durable data factory on top of a platform that's becoming the category standard, or we've handed a vendor significant leverage over our training pipeline. CVAT stays open-source and portable. Labelbox does not. That's the constraint you're accepting.

Hit 500 LBUs on Free and you can't add rows, labels or predictions until next month. Data stays exportable. Starter bills instead: card in the Billing tab, fixed $0.10 per LBU, no user limit.

Annotate charges 1 LBU per labeled row, once. Model charges 1 per 5 rows. Catalog charges 1 per 60 stored rows, monthly. Price the docs' own case: 60,000 images with Foundry predictions, 10,000 labeled. The docs bill 23,000 LBU in the first period, $2,300 including a month of storage, then $100 a month to keep the rows. Three years: $5,800, if the dataset stops growing. Rows move this bill. Seats don't.

The catch is everything off the meter. Foundry inference bills on top, by volume, model and task. SSO and HIPAA are Enterprise add-ons, requested through sales. I couldn't find prices for Horizon, Terra or Recursion. Enterprise is quote-only; its downgrades go through sales too.

Assets can stay in your own bucket. IAM delegated access lets Labelbox read straight from S3, Google Cloud Storage or Azure Blob, and Catalog layers vector and traditional search over 25+ cloud sources. Annotate carries 10+ editors, computer vision through LLM evaluation, and model-assisted labeling runs from the SDK as well as the app.

The gauge a team lead watches is the LBU meter. A labeled row costs 1 LBU once, every 60 stored rows cost 1 a month, and skipped rows count as labeled, so a messy queue shows up in usage. Starter unlocks from the Billing tab with no user or project cap and 5,000 API requests a minute against Free's 1,500. The catch is Free's hard stop: past 500 monthly LBUs you can export, not add rows or labels.

One oddity: labelbox.com's product menu lists Recursion, Horizon and Terra, while the labeling app is documented at docs.labelbox.com. Those docs read like operators wrote them. Model runs cap at 20 below Enterprise, though.

The product has clearly pivoted hard toward the serious AI lab buyer. The Alignerr Expert Network — 2.6M+ contributors, 50K+ PhDs, 40+ countries — isn't a freelancer pool, it's infrastructure. The reinforcement learning data features, Arena Evals, and Private AGI Benchmarks tell you exactly who this is built for: teams post-training foundation models, not someone labeling product photos. That's a real identity. Focused beats broad most days.

The free tier caps at 30 users and 50 projects, which is generous for evaluation. Past that, Starter is self-serve at a fixed $0.10 per LBU with no user limit. The catch for daily users is metering: labeled rows, stored rows and model runs all burn LBUs, so turning that rate into a monthly bill means working through the docs' consumption tables.

The web-only platform and zero mobile parity makes sense for annotation-heavy workflows, but it's worth naming. This is a desktop job, full stop. Onboarding a new ML engineer will take more than ten minutes — the feature surface is wide and the enterprise framing doesn't hold your hand much. Month three, though? Probably feels like home.

Three tells worth naming. One: the only published price — a fixed $0.10 per LBU on the Starter plan — sits in the docs rather than on a pricing page, and Horizon, Terra and Recursion are quoted through sales. Two: the Alignerr headcounts — 2.6M contributors, 50K PhDs — are vendor-stated, with nothing independent behind them. Three: the homepage now leads with RL environments and Recursion, a managed-agent platform, while the labeling app underneath still ships regular release notes. The old identity and new reality are out of sync.

The differentiation is real, though. Alignerr's 50K+ PhDs and 85K+ licensed professionals isn't something Scale AI or CVAT can replicate overnight. Private AGI benchmarks and Arena Evals put Labelbox inside the model development loop — not just upstream of it. That's a stickier position than annotation-only competitors.

Exit portability is the honest concern. Deep workflow integration, managed labeling services, and custom evals create lock-in by design. API access exists, but the Alignerr expert network isn't portable anywhere. If direction shifts, migration is painful.