Robot foundation models are trained on video, and video is useless to them until someone labels it: where each action starts, what the action was, and whether it worked. Nvidia's GEAR lab has now released the system it uses for that job. Flora, publicly presented on October 20, is three models fine-tuned from Alibaba's Qwen3.5 family. A 9B model, Flora-TempSeg, cuts a recording into action intervals. A 27B model, Flora-Captioner, writes one instruction for each interval, using the video before and after it to decide what the hands were actually doing. A second 27B model, Flora-Rewarder, grades the result: a success score for the whole episode, a normal, suboptimal or failure label for each half-second step, and scores against rubrics the user defines.
Nvidia says this is the data engine behind its GR00T robot models, running at a scale of 200,000 hours of labeled video a day across 512 GB300 GPUs. Treat that as the company's production claim rather than a verified figure. What can be checked today is the comparison Nvidia chose to headline, and it deserves a slower read than the multiple it produced.
The 125x is a projection set against an API price
Nvidia's research page describes Flora as "roughly ~125× cheaper than GPT-6 Astra". The numbers underneath, reported by The Decoder from Nvidia's own chart on October 25, are $0.72 per hour of source video for Flora and $88.49 for GPT-6 Astra. In a separate Nvidia-run comparison of 50 videos, Flora reportedly scored 57.1% against 53.7% for GPT-6 Astra at full reasoning effort. That result does not mean Flora leads across Nvidia's full evaluation suite. Using the displayed dollar amounts, ALFW calculates a ratio of approximately 122.9×, consistent with Nvidia's rounded claim of roughly 125×. The meaningful uncertainty is in the cost assumptions, not the rounding.
Nvidia's comparison, as reported by The Decoder, estimates Flora's compute cost using a $10-per-GPU-hour rental assumption and compares it with an estimated GPT-6 Astra API charge under the tested workload. These are not two independently observed production bills, and the calculation does not establish the total cost of running an annotation service. Hardware utilization, batching, preprocessing, storage, engineering and operating overhead may all change a buyer's actual cost, and change the rental assumption and the multiple moves with it. That does not make the direction wrong. A specialised 27B model running on a lab's own hardware will almost certainly label video for less than a frontier API at maximum effort. It makes the precision assumed, and precision is what a number like 125x claims.
Nvidia's own table does not give Flora the sweep
Nvidia publishes a detailed benchmark table that includes metrics on which Astra outperforms Flora. On human recordings in Nvidia's DenseCaptionBench, Flora leads clearly, with a joint point F1 of 37.4 against Astra's 29.7. On robot recordings from the same benchmark, the order reverses: Astra 57.9, Flora 51.0. On the public WGO-Bench, Astra leads the joint semantic F1 39.0 to 34.5 and gold-window accuracy 77.4 to 64.7, while Flora leads segmentation [email protected], 50.3 to 46.0. On rubric scoring, Astra's pooled L1 error is 3.6 against Flora's 6.5. On RoboReward, the order flips again: Flora's MAE is 40.0 against Astra's 55.4.
Flora's strongest result is the one with a practical use attached. Its rewarder caught 53.3% of failed steps inside an episode; Astra caught none, and Flora also leads DenseRewardBench macro F1, 46.2 to 40.7, while Astra leads the Robometer out-of-distribution Kendall measure, 82.3 to 72.0. The trade-off is recorded in the same research: Flora also flags about 11% of normal steps as failures. A monitor that misses half of all failures and cries wolf on one normal step in nine is a triage tool, not a judge. One more caveat is the researchers' own: on parts of the captioning evaluation, GPT-6 Astra grades the descriptions while also competing as a system under test. That is an evaluation-design limitation, not proof of bias, and none of these results is an independent validation of either system.
The license decides who gets the saving
The Hugging Face cards are the part of this release that will matter longest. Flora's released weights are governed by Nvidia's noncommercial license, and its model cards explicitly exclude commercial and production use from their intended applications. The cards describe the models as "for research and development only", state that the intended use is noncommercial robotics research, and say directly that Flora "is not intended or supported for commercial or production use". Commercial deployment cannot be assumed permissible under the published terms; organizations would need appropriate authorization or legal review. The Apache 2.0 license visible on the same page belongs to the Qwen3.5 base models, not to Nvidia's fine-tuned weights.
Nvidia's own use of Flora in production does not grant outside organizations commercial production rights under the downloadable license. So the 125x saving, even taken at face value, is not a price a robotics company can plan around. It is an offer to qualifying noncommercial research: the labeling pipeline behind GR00T, free to download, free to evaluate, and closed to the production budgets its cost chart appears to address. That is still a substantial offer. Labeled manipulation video is a bottleneck every robotics lab shares, and a downloadable segmenter, captioner and rewarder from the team training GR00T will be used. Just read the release for what it is: a research release with a production-scale claim on the first page and a research-only license on the download page. Both are Nvidia's own words. They point in different directions, and the second one is legally binding.
Sources
- [1] Nvidia Research, GEAR lab (October 20, 2026) — Flora project page: the three-label framing, the "roughly ~125× cheaper than GPT-6 Astra" claim, the 200,000-hours-per-day across 512 GB300s production claim, and the full benchmark table (DenseCaptionBench Human100 and Robot200, WGO-Bench, RoboReward, Robometer, DenseRewardBench, RubricBench) read first-hand by ALFWRead source
- [2] Hugging Face, nvidia/Flora-Captioner model card — model description, Qwen3.5-27B base, NVIDIA OneWay Noncommercial License, "research and development only", noncommercial intended use, commercial/production use not supported, and the October 15, 2026 Hugging Face release dateRead source
- [3] Hugging Face, nvidia/Flora-Rewarder model card — outcome and process reward description, the same noncommercial licence and use restrictions, and the October 15, 2026 release dateRead source
- [4] The Decoder (October 25, 2026) — the $0.72 vs $88.49 per-hour figures, the 57.1% vs 53.7% quality scores in the 50-video test, Claude Opus 5 at about $42 per hour, the statement that Flora's costs are projections based on an assumed $10-per-GPU-hour rental price, the 53.3% failure-recall and ~11% false-flag figures, and the note that Astra both competes in and helps grade the captioning evaluationRead source