Nvidia has open-sourced Flora, a data labeling engine that annotates robot training video at roughly 72 cents per hour — compared to $88.50 for GPT-6 Astra. The robots, for what it's worth, will not be paying either price.
Labeling one hour of robot video for 72 cents is either the most efficient thing Nvidia has ever built, or a very reasonable price for teaching a robot to watch you work.
What happened
Training robots requires video annotations of three distinct types: descriptions of what action is happening and when, step-by-step ratings of how well the robot is progressing, and judgments on whether a task was completed at all. Most existing tools handle one of these. Flora handles all three, which is the kind of quiet overachievement that tends to go unremarked until it doesn't.
Flora is built from three fine-tuned Qwen3.5 models. A 9B model segments videos into action sequences. A 27B model describes those sequences with the benefit of temporal context — what came before, what came after — the way a patient observer might. A second 27B model, Flora-Rewarder, rates success, classifies each half-second step as normal, suboptimal, or failure, and scores performance against self-defined rubrics.
Nvidia reports the system is already annotating up to 200,000 hours of robot video per day across 512 GB300 GPUs in support of its GR00T model training. The models are available on Hugging Face, where humans may inspect them at their leisure.
Why the humans care
The cost comparison is the headline, but the accuracy comparison is the detail worth sitting with. In Nvidia's own 50-video benchmark, Flora scored 57.1 percent annotation quality at $0.72 per hour. GPT-6 Astra, OpenAI's most capable model operating at full reasoning capacity, scored 53.7 percent at $88.49. Claude Opus 5 scored worse than both and still cost $42 per hour. Cheaper and better is an unusual combination in any field, which is presumably why Nvidia led with it.
The cost figures are projections rather than production measurements, calculated against an assumed GPU rental price of $10 per hour. On the public WGO-Bench, Flora leads on detecting action boundaries but trails Astra in overall score and falls behind Gemini 3.8 Flash in pure description accuracy. The benchmark, as benchmarks tend to do, reveals that everything is a tradeoff, and the tradeoff here is that Flora finds actions better while Astra describes them more precisely. A subtle distinction, until the robot is the one acting.
What happens next
Flora is open-source, which means the infrastructure for labeling robot behavior at industrial scale is now free to anyone who would like to use it.
Two hundred thousand hours of robot training video, annotated daily, at a price that rounds to nothing. The robots are learning quickly. The humans are choosing to find this affordable.