Flora, Nvidia's robot video labeler, was priced at $0.72 an hour against $88.49 for GPT-6 Astra
Nvidia's GEAR lab says its three-model suite scored 57.1% to Astra's 53.7% on a 50-episode subset, but the cost comparison is not like-for-like.
In brief
Nvidia's Flora labels robot video for $0.72 an hour
Nvidia estimates Flora costs about $0.72 per video hour against $88.49 for GPT-6 Astra. That is roughly 125x cheaper, but the comparison is not like-for-like.
Nvidia's GEAR lab has released Flora, a set of three AI models that cut robot video into steps, describe what happens, and score how well the robot did. We know this from the lab's research page, its paper and the Hugging Face model card. The release is dated October 20, 2026.
What is in the box
Flora has one 9B segmenter and two 27B models. All three are built on Qwen3.5. Together, the suite segments robot video, writes captions and gives scores.
The numbers
Nvidia's paper estimates about $0.72 per hour of source video. It puts GPT-6 Astra at $88.49. That is a ratio of roughly 122.5, which Nvidia rounds to about 125x.
On a 50-episode subset, Flora scored 57.1% and Astra 53.7%. On RoboRewardBench, the paper reports a 27.8% lower error than Astra.
Nvidia also reports labeling 200,000 hours of video per day on 512 GB300 GPUs. That is Nvidia's own production claim. Nothing in the sources confirms it independently.
Why it matters
Robots learn from video, and someone has to label it first. Imagine a million hours of footage where every clip needs a note: what was tried, and did it work? Doing that by hand is slow. Doing it through a paid API gets expensive fast. A cheaper labeler makes bigger robot datasets more affordable.
Read the cost claim carefully
The two prices are not measured the same way. Flora's figure is a modeled projection that assumes $10 per GPU-hour. Astra's figure comes from actual API usage. Coverage of the release flags that this is not a like-for-like comparison.
Not a clean sweep
Nvidia's paper says Flora is stronger on human video and on cases of partial success. Astra leads on robot video, caption accuracy and rubric scoring. Only some comparisons are reported.
Open, with strings
Nvidia's page calls the suite open. The model cards, though, specify a noncommercial OneWay license, limited to research and evaluation. Companies cannot simply drop it into a commercial product.
The limits: the cost figure is a projection, the accuracy test used 50 episodes, and the daily labeling volume is unverified. Astra still wins on robot video, the very footage robots produce.