Flora labels 200,000 hours of data per day across 512 GB300s. That number does not prove a measured cost per video hour, an energy draw, or a license that lets a buyer reproduce the run. The GEAR page pairs the scale claim with a roughly 125x price gap against GPT-6 Astra, but the gap is a vendor comparison, not a production invoice.

API label buyers are the constrained side. If a robotics lab can rent or own GB300s, the open suite shifts spend from per-token pricing to GPU-hours and gives it outcome, process, and caption labels in one pipeline. Benchmark rows still leave Astra ahead on WGO-Bench Gold-window accuracy and joint semantic [email protected], so the suite is not a blanket quality win.

No production cost, energy draw, license terms, cumulative hours, or overall WGO-Bench score appears in the document yet.

The 200,000-hour figure proves a large internal labeling pipeline, not a buyer-ready price. It changes the cost model for labs that can run GB300s, while leaving API-dependent teams exposed to per-token pricing.

The document leaves out measured production cost, energy draw, license terms, cumulative hours, and an overall WGO-Bench score.

Flora turns label generation into a GPU-hour workload on 512 GB300s, so the economic constraint moves from API tokens to cluster availability. The benchmark table still shows Astra ahead on WGO-Bench Gold-window accuracy and joint semantic [email protected], so quality is not uniform.