Nvidia's open Flora models label robot video at 125x less than GPT-6 Astra

Nvidia's GEAR research lab published Flora on its project page, and The Decoder reported it on 25 October 2026. Flora is an open set of models that labels video of robots and people doing tasks. Nvidia says it is about 125 times cheaper than labeling the same video with GPT-6 Astra, and that it is competitive on quality. The saving is Nvidia's projection, and its own results table shows where the cheaper models lose.

What Flora does

Nvidia says robot foundation models need three kinds of label. Timestamped captions are used to filter and balance pretraining data. Process rewards are used during reinforcement-learning post-training. Outcome rewards are used to monitor deployed robots. Nvidia says Flora covers all three and that it is the data engine used to train its GR00T models.

According to The Decoder's reading of the paper, Flora is three fine-tuned Qwen3.5 models: a 9B model that splits video into action segments, a 27B model that describes them, and a 27B model called Flora-Rewarder that scores an episode from 1 to 5 and rates each half-second step as normal, suboptimal or failure. The Hugging Face collection lists Flora-Captioner, Flora-TempSeg and Flora-Rewarder. I did not check the licence terms.

The cost claim

Nvidia says Flora labels data at roughly 125 times lower cost than GPT-6 Astra and is running in production at 200,000 hours of data per day on 512 GB300 GPUs. The Decoder reports the figures behind the ratio: about $0.72 per video hour for Flora against $88.49 for Astra, and 57.1 percent against 53.7 percent annotation quality on a 50-video test with Astra at full reasoning effort. The Decoder notes that the Flora cost is a projection from an assumed $10 per GPU hour, not a production measurement.

Where the cheaper models lose

Nvidia's table is more mixed than the headline. On human-recorded video, Flora scores ahead of Astra on caption segmentation (54.6 against 44.2 F1@0.5). On robot video, Astra leads: 71.9 against 66.9 on segmentation and 80.2 against 73.7 on caption point quality. On the public WGO-Bench, Astra leads on gold-window accuracy (77.4 against 64.7). On RubricBench, where lower is better, Astra's pooled error is 3.6 against Flora's 6.5.

Flora's clearest advantage is failure detection. On DenseRewardBench, Nvidia reports failure recall of 53.3 for Flora against 0.0 for Astra. The Decoder also flags a method issue the researchers acknowledge: GPT-6 Astra helps evaluate descriptions while also being one of the models under test. Most benchmarks here are Nvidia's own or public robotics sets, so none of them is an independent test.

What to take from it for other labeling jobs

• If you pay a frontier API to label or grade large volumes of one data type, price a tuned open model against it, and include GPU rental in the comparison as Nvidia did.
• Check the rows that match your task, not the average: Nvidia's own table has the cheaper model behind on some robot-video measures and well ahead on failure recall.
• Treat a projected saving as a hypothesis until you have run it on your hardware; Nvidia assumed $10 per GPU hour.
• Do not let the model you are testing also grade the test, which is the issue the researchers themselves acknowledge.

Unknown from the sources I read: Flora's licence, the hardware needed to run it on a small scale, and any result from outside Nvidia.