# Nvidia open-sources Flora, cutting robot video labeling costs 125x versus GPT-6 Astra

> Nvidia released the three-model Flora suite for robotics data labeling under a non-commercial license, while Anthropic's Applied AI team reported that consolidated single agents now run roughly twice as fast as multi-agent systems.

*Published 2026-09-28 · Updated 2026-10-25 · By The Intel Desk*

Flora is an open-source suite of three fine-tuned Qwen3.5 models — 9B TempSeg, 27B Captioner and 27B Rewarder — released by Nvidia to generate timestamped action captions, process rewards, outcome rewards and rubric scoring from robot and human video for training Nvidia's GR00T robot foundation models.

Nvidia's release turns a proprietary labeling pipeline into an open research asset. Flora is available on Hugging Face under the NVIDIA non-commercial license, a licensing choice that lets academic and corporate research teams reuse the weights while restricting commercial redistribution. The suite is built around three fine-tuned Qwen3.5 models, and Nvidia describes it as an industrial-strength data engine used for training GR00T models, according to the Flora research page.

Flora's cost advantage is the centerpiece of the announcement. Nvidia reports that labeling one hour of robot video costs about $0.72 with Flora, compared with roughly $88.50 with GPT-6 Astra, a roughly 125x reduction. The company says Flora is battle tested at scale, labeling 200,000 hours of data per day across 512 GB300 GPUs.

Jim Fan, Nvidia director of robotics and distinguished scientist, framed the release as a direct challenge to commercial API alternatives. 'We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today!' Fan wrote on X. The post positions Flora as a competitive alternative rather than a research toy.

> We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today!Jim Fan, NVIDIA Director of Robotics & Distinguished Scientist

## Why is robotics video labeling a bottleneck?

Robot learning depends on large volumes of labeled video, but labeling is expensive and slow. Human annotators must watch footage and write timestamped descriptions of actions, and their output is inconsistent. API-based models can automate the task, yet per-hour pricing makes large-scale labeling uneconomical, which is the problem Flora targets.

Video also carries more information than static images. A robot needs to know not just what object is present but when a grasp begins, when a joint reaches a target, and whether an action succeeded. That temporal structure is exactly what Flora's TempSeg and Captioner models are designed to extract.

Robot foundation models are trained to map sensor observations to actions, and the quality of that mapping depends on the diversity and volume of demonstration video. A model that has only seen a few hundred hours of picking-and-placing video will fail on new objects, new lighting and new arm configurations. Scaling the dataset is therefore the fastest way to improve generalization.

Labeled video is also the substrate for reward modeling. Instead of hand-designing reward functions for every task, researchers can learn rewards from video demonstrations, which is what Flora's rewarder does. This removes a major bottleneck in reinforcement learning for robots, where reward specification has historically been manual and brittle.

The economics of that data determine how many labs can participate. When labeling costs $88.50 per hour, only well-funded organizations can assemble large datasets. When the same hour costs $0.72, mid-sized labs and universities can enter the field, which expands the volume and variety of training data available to the whole community.

## What is Flora and why did Nvidia open-source it?

Flora addresses the labeling bottleneck by splitting the work into three specialist models. The 9B TempSeg model identifies temporal boundaries in video, the 27B Captioner produces timestamped action captions, and the 27B Rewarder scores each segment with process rewards, outcome rewards and rubric scoring. Nvidia designed the suite to work on both robot and human video, widening the pool of usable training data beyond teleoperation logs.

Open-sourcing the weights is a strategic departure from the API model. Teams can run Flora on their own GPUs, avoid per-hour labeling fees, and fine-tune the models further. The non-commercial license limits what companies can do with the weights, but it still removes the marginal cost of every API call, which is the main driver of the 125x gap.

Open weights also encourage adoption by the research community, which is Nvidia's core audience for GR00T. The more teams build data pipelines around Flora, the more training data and model improvements flow back into the ecosystem. Nvidia's announcement frames the release as a contribution to the field rather than a product launch.

## How much cheaper is Flora than GPT-6 Astra?

The per-hour cost figures come from Nvidia's Flora research page and were also reported by The Decoder. Nvidia's paper states the model is substantially cheaper than API-based alternatives, roughly 125x cheaper than GPT-6 Astra. The Decoder put the comparison in dollar terms: about $0.72 per hour with Flora versus roughly $88.50 with GPT-6 Astra.

The cost gap reflects architectural and business-model differences. GPT-6 Astra is a commercial API product whose pricing includes serving, margin and overhead. Flora's open weights run on the customer's own infrastructure, and Nvidia can amortize the 512 GB300 GPUs used for its internal labeling pipeline. Once weights are downloaded, additional video hours cost only compute.

Scale matters because robot data is scarce. A fleet of robots can generate far more video than a human annotation team can label, and API-based labeling at $88.50 per hour quickly becomes uneconomical. At the per-hour prices Nvidia reports, labeling 200,000 hours per day would cost roughly $17.7 million with GPT-6 Astra pricing and about $144,000 with Flora.

| Metric | Flora | GPT-6 Astra |
| --- | --- | --- |
| Models | Three Qwen3.5 fine-tunes (9B, 27B, 27B) | Commercial API, model size not disclosed |
| Per-hour labeling cost | $0.72 | Roughly $88.50 |
| Cost ratio | 1x | ~125x more expensive |
| Internal 50-video annotation quality | 57.1% | 53.7% |
| Throughput | Up to 200,000 hours/day across 512 GB300 GPUs | Not disclosed |
| License | NVIDIA non-commercial, open weights | Proprietary API |

## How does Flora compare with GPT-6 Astra on quality?

Cost is only half of the comparison. On an internal 50-video test reported by The Decoder, Flora scored 57.1% annotation quality versus 53.7% for GPT-6 Astra, a 3.4 percentage point advantage. Nvidia says Flora is competitive with or beats GPT-6 Astra on annotation quality while being substantially cheaper.

The benchmark is internal, so it is not an independent apples-to-apples evaluation. Still, the result undermines the assumption that frontier API models are necessarily the best labelers. Fine-tuned specialist models trained on robotics data can outperform a general-purpose model on this narrow task, especially when the task rewards timestamp precision and domain-specific language.

Quality also depends on downstream use. Flora's rewarder outputs are designed for GR00T training, which means the suite is optimized for Nvidia's own model family rather than for every possible robotics stack. Teams using different base models may need to adapt the captions or rewards, and the non-commercial license limits how much value they can extract commercially.

## How does Flora work technically?

Flora's pipeline is a three-stage process. TempSeg first splits a video into temporally bounded segments; Captioner then writes a natural-language action caption for each segment with timestamps; Rewarder finally scores each segment using process rewards, outcome rewards and rubric scoring. The output is a labeled dataset plus reward signals that can be used directly in GR00T training.

- TempSeg (9B) segments raw video into temporally bounded action units.
- Captioner (27B) generates timestamped natural-language captions for each segment.
- Rewarder (27B) assigns process rewards, outcome rewards and rubric scores.
- The combined labels and rewards are used to train GR00T robot foundation models.
- Developers can filter or prioritize training clips using the rubric scores.

Each model is a fine-tune of Qwen3.5, which means Flora inherits the base model's instruction-following and generalization behavior while adding robotics-specific skills. The 27B rewarder is the component that most directly affects learning: process rewards give credit for correct sub-steps, outcome rewards signal final success, and rubric scores let developers set task-specific thresholds.

Running the pipeline at scale requires serious infrastructure. Nvidia reports labeling up to 200,000 hours of video per day across 512 GB300 GPUs, which is far beyond what a single research lab could produce. That throughput figure is a strong signal that Flora is not a demo; it is the production engine behind GR00T training runs.

## What does Anthropic's single-agent finding mean?

In a separate development reported by The Decoder, Anthropic's Applied AI team found that consolidated single-agent systems run roughly twice as fast as multi-agent setups. The finding is notable because multi-agent orchestration has become a common pattern for complex tasks, with developers splitting work among subagents that each handle a narrow function.

Anthropic's explanation is that modern models can manage complex context and tasks internally, so the overhead of coordinating subagents — message passing, duplicated context, scheduling — outweighs the benefit. In other words, the model has become the orchestrator, and external orchestration layers add latency without adding capability.

The two stories point in the same direction: consolidation. Flora consolidates three specialist functions into a single open-source pipeline, and Anthropic's data suggests agent systems should consolidate subagents into the main loop. For engineering teams, the practical takeaway is to measure overhead before adding architectural complexity.

## What are the market and stakeholder implications?

For robotics startups, the most direct effect is a lower cost of entry. Using Nvidia's stated per-hour cost, a team can label 100,000 hours of video for $72,000, a sum that was previously impractical for most labs. That changes which research projects are fundable and which datasets are worth collecting.

For Nvidia, Flora is an ecosystem play. Cheaper labeling means more teams can train robot foundation models, and many of those teams will run training on Nvidia hardware. The GB300 GPUs used in Flora's own pipeline also become a reference deployment that Nvidia can point to when selling data-center systems.

For commercial API providers such as GPT-6 Astra, open-source alternatives create pricing pressure. If a fine-tuned open model can match or beat a frontier API on a specific task, the API's premium becomes harder to justify. The robotics labeling market may follow the pattern seen in text generation, where open-weight models compress prices over time.

For the broader AI market, the Anthropic finding could reshape agent frameworks. If single agents are twice as fast, product teams may simplify their orchestration layers, reduce token spend and improve responsiveness. The shift would favor model-level context management over external subagent coordination.

## What comes next?

Flora's release is likely to be followed by community adaptations, including fine-tunes for specific robot morphologies and integrations with popular robotics stacks. The non-commercial license will push commercial users toward Nvidia's own offerings or custom licensing, which may be part of the strategy.

On the research side, the 57.1% versus 53.7% quality result is a single internal benchmark, and independent evaluations will be needed to confirm the advantage. The open weights make such evaluations possible, because any lab can run Flora on its own test videos and compare outputs.

Anthropic's single-agent finding will likely be tested in production systems over the coming quarters. The twice-as-fast claim is specific to the architectures Anthropic studied, and the right pattern may still depend on task complexity, model size and latency requirements. The general direction, however, is consistent with a maturing field that is moving away from unnecessary orchestration.

Together, the two developments signal that the frontier of applied AI is shifting from raw capability to efficiency. Nvidia is attacking the cost side of robotics data, and Anthropic is attacking the overhead side of agent systems. Both are attempts to make AI systems cheaper and faster to operate at scale.

## Sources

1. [Flora is an industrial-strength data engine used for training GR00T models, is roughly 125x cheaper than GPT-6 Astra, and labels 200,000 hours of data per day across 512 GB300s.](https://research.nvidia.com/labs/gear/flora/)
2. [Flora is substantially cheaper than API-based alternatives, roughly 125x cheaper than GPT-6 Astra, and is battle tested at scale.](https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf)
3. [Flora labels one hour of robot video for about $0.72 versus roughly $88.50 with GPT-6 Astra, and scored 57.1% versus 53.7% on an internal 50-video annotation quality test.](https://the-decoder.com/nvidia-open-sources-flora-to-make-robot-video-labeling-125x-cheaper-than-gpt-6-astra/)
4. [Anthropic's Applied AI team reported that a single-agent architecture ran about twice as fast as a multi-agent setup.](https://the-decoder.com/anthropic-says-one-ai-agent-beats-many-subagents-now-that-models-can-handle-complexity-alone/)
5. [Anthropic's Applied AI team published findings on when to fold subagents into the main agent loop.](https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop)
6. [Jim Fan announced Flora's open-source release and called it a robotics data engine with Astra-level performance at 125x lower cost.](https://x.com/DrJimFan/status/2112597587987689912)
7. [Anthropic reported that consolidated single agents now run roughly twice as fast as multi-agent systems because modern models can manage complex context and tasks internally.](https://aibroadsheet.com/news)

---
Source: https://aiintelreport.com/frontier-models/frontier-models
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt