🔍 Read the full analysis: The 125× Cost Claim For Nvidia Flora’s Robot Video Labeling, Explained on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, describing and scoring robot video. Its paper estimates captioning costs at $0.72 per video hour, about 122.5 times below GPT-6 Astra’s $88.49 API cost on a 50-episode test—but Flora’s cost is modeled, and its license limits use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, three models designed to label video used in robot training, and estimates their captioning cost at $0.72 per video hour compared with $88.49 for GPT-6 Astra. The paper describes the gap as approximately 125 times, but the comparison is not like-for-like: Flora’s figure is a modeled GPU cost, while Astra’s is measured API usage on 50 test episodes.
Flora combines three models built on Alibaba’s Qwen3.5 family. A 9-billion-parameter segmenter identifies when actions begin and end; a 27-billion-parameter captioner writes an instruction for each segment; and a second 27-billion-parameter model scores a full attempt or individual steps. These labels can help robotics researchers organize training video and identify successful, suboptimal or failed actions.
The paper’s cost test used 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra, run at maximum reasoning effort, incurred $88.49 per video hour at public API rates listed for October 4, 2026. Nvidia modeled Flora at $0.72 per hour using eight GB300 GPUs and an assumed price of $10 per GPU-hour. The paper gives a range of $0.51 to $1.26, excluding storage and other overhead. Dividing $88.49 by $0.72 gives 122.5, the basis for the rounded 125-times claim.
On the same test episodes, Nvidia reported an overall score of 57.1% for Flora and 53.7% for Astra. Results varied by task: Flora scored better on human-hand videos, action boundaries and some partial-success cases, while Astra led on robot videos, caption accuracy and rubric scoring. The cost headline concerns captioning; it does not include the reward model.
Infographic 01 · What Flora does
Three models that turn robot video into training labels.
Flora cuts a recording into actions, describes each one and grades how well it went. The 200,000-hour figure is Nvidia’s own production claim.
Flora-TempSeg
9B · adapter on Qwen3.5-9B
- Job: Finds where each action starts and ends
- In: Video windows (one camera, or main + two wrist cameras)
- Out: Time intervals as JSON
Flora-Captioner
27B · fine-tuned Qwen3.5-27B
- Job: Describes each action, using the video before and after
- In: One interval plus context
- Out: One English instruction, e.g. “Place the blue mug on the tray with your right hand.”
Flora-Rewarder
27B · fine-tuned Qwen3.5-27B
- Job: Judges how well the robot did
- In: An episode or a 16-second window, plus the task
- Out: Score 1–5 per episode · normal / suboptimal / failure per 0.5 s · rubric scores
Three labels, three uses
Why robot models need each one, per the paper
| Label | Used for | Made by |
|---|---|---|
| Timestamped captions | Filtering and rebalancing pretraining data | TempSeg + Captioner |
| Process rewards | Reinforcement learning after pretraining: which step went wrong | Rewarder |
| Outcome rewards | Monitoring robots after deployment | Rewarder |
What it is not
model cards
- Not a robot controller: outputs are annotations only
- Not meant to be the sole safety monitor for a physical system
- Not for commercial or production use (noncommercial license)
- Hardware: tested on Nvidia H100, Linux; ~55 GB of weights per 27B model (estimate)
Sources: Flora paper, Nvidia (20 Oct 2026); Hugging Face model cards (15 Oct 2026); Nvidia GEAR project page. Weight-size estimate: Thorsten Meyer (27B parameters × 2 bytes).
Powered by Thorsten Meyer AI
What Cheaper Robot Labels Could Change
Video annotation is one step in turning demonstrations into data that robot models can use. Lower labeling costs, if reproduced at scale, could let robotics teams process more examples or spend less on an existing workflow. That matters because training data needs to capture not only successful motions but also mistakes, interrupted tasks and differences in how people or machines perform an action.
The reported saving is not a demonstrated operating cost for companies. Nvidia says its Flora models supported data preparation for GR00T-2, but the paper’s benchmark does not establish the cost of running a full production pipeline. Flora’s license also limits use to research or evaluation, making it a research resource and comparison point rather than a free commercial labeling service. Data collection, robot hardware and safety remain separate constraints.
Popular right now

De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Milk Frother, Silver

Breville BES870BSXL Barista Express Espresso Machine, Black Sesame

Vitamix 5200 Blender with 64 oz Container, Black
As an affiliate, we earn on qualifying purchases.
How Flora Labels Robot Footage
Robot foundation models learn from recordings of tasks such as placing objects, folding fabric or manipulating tools. The paper describes three label types: timestamped captions that connect an instruction to a video interval, process rewards that grade short steps, and outcome rewards that score an entire attempt. The Flora models address those tasks separately rather than packaging every function into one model.
According to the model cards, Flora was trained on fewer than 10,000 hours of human and robot video. The models are available through Hugging Face repositories, with inference code included. Nvidia says its system can process up to 200,000 hours of video per day using 512 GB300 GPUs; that is a company-reported capacity figure, not an independently verified production result. The stated license is the NVIDIA OneWay Noncommercial License, restricted to research or evaluation use, with derivatives subject to the same limit.
“The production GPU-hour denominator has not been verified.”
— Nvidia, in the Flora paper’s cost discussion
What the Cost Test Cannot Establish
The 125-times figure is a comparison of different kinds of costs: a modeled estimate for Flora and a measured API subtotal for Astra. Nvidia’s test also covered a limited set selected from its own benchmark, not a broad survey of production workloads. The paper reports that 106 of 7,062 Astra attempts had no usage data, so the $88.49 figure is a known-usage subtotal. Nvidia says the GPU-hour denominator behind Flora’s estimate has not been verified.
The headline does not show whether the same cost gap holds with other GPU prices, utilization rates, video formats or workloads. It also does not include storage and overhead in Flora’s estimate, and the 125-times comparison covers captioning rather than every Flora model. Nvidia ran most of the reported tests itself; independent replication and results on a wider range of videos are not established by the supplied material. The license’s application to specific commercial arrangements may also need clarification from Nvidia.
Replication and Licensing Questions
The next useful test is whether independent teams can reproduce Flora’s throughput, labeling quality and costs on larger and more varied robot datasets. A fuller comparison would report GPU utilization and all operating costs, evaluate the models across additional tasks, and separate captioning from segmentation and reward scoring.
For companies considering the models, the immediate next step is to check the license and obtain clarification from Nvidia before using Flora in a commercial workflow. The paper and model cards provide a starting point for research evaluation, but the materials cited here do not announce a revised license, independently verified production pricing or a schedule for further benchmark results.
Key Questions
What does Nvidia’s 125-times figure compare?
It compares Nvidia’s modeled Flora cost of $0.72 per video hour with a measured GPT-6 Astra API cost of $88.49 per video hour on 50 test episodes. The exact ratio is 122.5, rounded to approximately 125 times.
Was Flora’s $0.72 cost measured in production?
No. Nvidia calculated it using eight GB300 GPUs, an assumed price of $10 per GPU-hour and modeled throughput. The paper says the production GPU-hour denominator has not been verified; its range is $0.51 to $1.26 per video hour, excluding storage and overhead.
Did Flora outperform GPT-6 Astra on quality?
Nvidia reported a higher overall score for Flora on the selected test set, 57.1% versus 53.7%. Performance depended on the task: Flora led on some human-video and action-boundary tests, while Astra scored better on robot videos, caption accuracy and rubric scoring.
Can businesses use Flora commercially?
The supplied model cards list the NVIDIA OneWay Noncommercial License, for research or evaluation purposes only, with derivatives inheriting that limit. Organizations should review the license and seek clarification from Nvidia for a specific use.
Source: ThorstenMeyerAI.com
