Auto Label with GPT-6 Astra in Roboflow

Roboflow has made GPT-6 Astra, released in September 2026, the default Auto Label model for object detection projects, enabling zero-shot auto-labeling by typing class names to draw bounding boxes. The feature can be run from the Auto Label interface or via the Roboflow MCP server, allowing AI agents to handle the full computer vision pipeline. Roboflow Universe hosts over 1 million datasets and 250,000 models to support such workflows.

You can auto-label an object detection dataset in Roboflow by typing your class names and letting GPT-6 Astra, the default Auto Label model, draw the bounding boxes zero-shot. Run it from the Auto Label interface or hand the job to an agent through the Roboflow MCP server. How you label your data has a big impact on how quickly you get to your end goal: a working computer vision model. At Roboflow, we persistently strive toward reducing the amount of time you need to spend labeling so you can start using a model in production. Zero-shot detection https://blog.roboflow.com/what-is-zero-shot-object-detection/ is one of the biggest steps we have taken toward that goal. Instead of drawing each box by hand, you name the objects you want to find and a vision-language model locates them for you. In Roboflow, this runs through Auto Label https://docs.roboflow.com/datasets/annotate/annotate/ai-labeling/auto-label?ref=blog.roboflow.com , and its default model for object detection projects is now GPT-6 Astra https://playground.roboflow.com/models/openai/gpt-6-astra?ref=blog.roboflow.com , released in September 2026. Auto-labeling is just one piece of the workflow. With the Roboflow MCP server https://roboflow.com/mcp?ref=blog.roboflow.com , an AI agent can run the whole computer vision pipeline, from finding and labeling data to training and deployment. Roboflow becomes a specialized vision sub-agent that a frontier model can call on. In this guide, I'll walk through two ways to label a dataset with GPT-6 Astra: through the Auto Label interface, and by handing the job to an AI agent over the MCP server. Both paths end in the same place, a reviewed, training-ready dataset. Without further ado, let's get started. How GPT-6 Astra Labels Images in Roboflow With GPT-6 Astra, Auto Label can detect the classes you name and draw bounding boxes around them automatically. You do not need to train the model or provide example boxes, which makes it a true zero-shot workflow. Before you start, there are three things about how Astra labels images that can affect how you define classes and review the results. - Class names matter most. Astra uses the class name as the prompt, so keep it clear and specific. The description field does not affect the result. - There is no confidence filtering. Every detected box is returned, so you will need to remove incorrect labels during review. - Astra returns boxes, not masks. It handles object detection, not segmentation, so use another model if you need precise object shapes. Running GPT-6 Astra in Auto Label Step 1: Find a dataset on Roboflow Universe Roboflow Universe https://universe.roboflow.com/?ref=blog.roboflow.com hosts a large collection of open-source computer vision datasets, so you can often start from images someone has already gathered instead of collecting your own. It holds more than 1 million datasets and 250,000 models, which is usually enough to find something close to what you need. For this walkthrough, you will use the Construction Site Safety https://universe.roboflow.com/roboflow-universe-projects/construction-site-safety?ref=blog.roboflow.com dataset, published by Roboflow. It is an object detection set of work-site images showing people in protective equipment, which gives GPT-6 Astra clear, nameable objects to detect. You are only using this dataset as a source of images. You will not use its existing labels, so any object detection dataset with the objects you care about would work here. Step 2: Create an object detection project In Roboflow, open the Projects page and click New Project in the top right. Give the project a name here, safety-gear-labeling , leave visibility as Private, and set the project type to Object Detection since GPT-6 Astra returns bounding boxes. Then click Create Private Project. The project starts empty, ready for you to bring in the images in the next step. Step 3: Clone the raw images into your project Open the Construction Site Safety https://universe.roboflow.com/roboflow-universe-projects/construction-site-safety?ref=blog.roboflow.com dataset, select your images, then choose Clone Images and Clone Selected . This example uses 37 images for a quick batch test. In the Clone Images dialog, choose Import raw images only so Astra labels them from scratch. Then select your workspace and safety-gear-labeling project, and click Clone Images . Once the clone finishes, the images land in your project without labels, ready for GPT-6 Astra to annotate. Step 4: Open Auto Label on the cloned batch Go to your project and open the Annotate tab. The cloned batch appears under Unassigned. Click Annotate Images on the batch card. Label in minutes is selected by default on the right. From there, the Auto Label interface opens so you can confirm the model. GPT-6 Astra is already selected as the model, since it is the default. You will confirm it in a moment, but first you need to add the classes you want to detect. Step 5: Enter your classes In the Classes panel, add the classes you want Astra to detect and leave the description blank, since it only reads the class name. You can use whatever classes fit your task. In this example, the classes are hard hat, mask, and gloves. Add your classes before selecting the model, because choosing GPT-6 Astra runs a preview automatically, so the classes need to be in place first. With them set, you are ready to confirm the model. Step 6: Confirm GPT-6 Astra and preview the results GPT-6 Astra is already selected, so once your classes are added, Roboflow tests them on a few sample images and draws the boxes automatically. You can use Change to try different samples. Here, Astra correctly separates the hard hat, mask, and gloves into their own classes. Since all boxes return at full confidence, the shown and total counts are always the same. The second preview shows that Astra follows the class name, not the description. Even though hard hat has the description green vest, the model still detects hard hats. Once the preview looks good, run GPT-6 Astra on the full batch. Step 7: Label the full batch Click Auto Label With This Model in the top right to start labeling all selected images. Review the model, image count, and cost, then click Start Auto Label . For 37 images, the job costs 0.037 credits, plus separate OpenAI API usage. When the job finishes, the labeled images move to the Review column, where you can check and correct Astra’s results. Step 8: Review and fix the labels Open the batch from the Review column. Roboflow shows each image with GPT-6 Astra's boxes, the annotation tools down the right side, and Reject or Approve at the top. GPT-6 Astra labels only what it finds. In this tunnel image it detected hard hats but no gloves or masks, so those two classes sit under Unused Classes. The hard hats, though, are over-detected: the image came back with 13 boxes, several of them duplicates stacked on the same helmet, and a few drawn loosely around it. This is what the review pass is for. Using the tools on the right, you delete the duplicate boxes so each helmet keeps a single box, then drag the handles to tighten the loose ones around the hard hat. If a box has the wrong class, change it from the class list instead of deleting it, and add a box for any gear GPT-6 Astra missed. After this cleanup, the same image holds 8 clean hard hat boxes, one per helmet. When the image looks right, press Approve. If it needs more work, press Reject to send it back. Because every box comes back at full confidence, nothing is filtered out for you, so this pass is where the quality of the dataset is set. Step 9: Add the labeled images to your dataset After checking an image and making any corrections, click Approve. Work through the batch, approving each image once it is right. You can also use Approve All to accept the remaining images at once. Once every image is approved, click Add Approved To Dataset. The batch moves into your dataset, where the labeled images are collected together. From here the dataset is fully labeled: 37 reviewed images built from three class names and a quick correction pass, instead of drawing every box by hand. It is ready to version, train, or export. Labeling Through the MCP Server Everything so far happened in the Roboflow app. You can also hand the same job to an AI agent through the Roboflow MCP server https://docs.roboflow.com/agents/mcp-server?ref=blog.roboflow.com . MCP is an open standard that lets an agent call external tools, and Roboflow exposes its labeling actions as tools. You describe the job in plain language, the agent runs it, and the labeling still happens on Roboflow's side. It is a second way in, not a different result. Step 1: Connect Roboflow to your agent The server works with any MCP-compatible agent Claude, Cursor, or Codex and signs in with OAuth, so there is no API key to manage. For Claude, open the Roboflow connector https://claude.ai/directory/connectors/dbbc26cf-80b0-4a85-9877-f85874282794?ref=blog.roboflow.com , confirm, and sign in. It then works across Claude.ai, the desktop app, and Claude Code. Setup for the other agents is in the MCP server docs https://docs.roboflow.com/agents/mcp-server?ref=blog.roboflow.com . Step 2: Clone a fresh batch Start fresh with an empty project, then clone a small set of raw images into it. This gives the agent a clean, unlabeled batch to work with. Step 3: Run the job from a prompt Ask the agent to label the batch: Label the new batch in my safety-gear-labeling project with Auto Label using GPT-6 Astra, with the classes hard hat, mask, and gloves. The first time the agent reaches into your workspace, it asks permission before running anything. You approve the call and it continues. The agent finds the 13-image batch and starts Auto Label using GPT-6 Astra with the three selected classes, while also tracking the job and reporting the project details. You can name Astra in the prompt, but you do not have to if it is already the workspace default. Step 4: Pick up the results in Review Once the job finishes, all 13 labeled images appear in the Review column, where you can open the batch, fix any duplicates or missed labels, and approve the images into your dataset. Both paths produce the same reviewed dataset. The MCP route just takes the mechanical part off your hands. Where GPT-6 Astra Fits Among Auto Label Models When it comes to auto-labeling, different AI models bring different strengths to the table. GPT-6 Astra stands out for its ability to understand context and handle a wide range of labeling tasks. - What sets GPT-6 Astra apart: GPT-6 Astra focuses on understanding context, relationships, and detailed instructions. This makes it useful for complex datasets where labels require more reasoning instead of simple object detection. - How it compares to Gemini and SAM 3: Gemini is strong at working with text and visual information together, while SAM 3 is more specialized in image segmentation and identifying precise objects or regions. GPT-6 Astra is better suited to tasks where deeper contextual understanding matters. - Choosing the right model: The best model depends on the type of data, the level of detail required, and the labeling workflow. Overall, GPT-6 Astra fits well into workflows that need more than basic detection. While Gemini and SAM 3 may be better for certain visual or multimodal tasks, Astra can be a strong choice when accurate labeling depends on understanding the bigger picture. Conclusion Roboflow gives you two simple ways to auto-label with GPT-6 Astra: through the Auto Label interface or with an agent over MCP. The interface is best when you want to control each setting, while the agent is useful for speeding up the setup. In both cases, Astra labels from class names, returns bounding boxes, and may still need cleanup for duplicates, loose boxes, or missed objects. After reviewing and correcting the results, your batch is ready to be turned into a dataset version and used for model training. Further reading Cite this Post Use the following entry to cite this post in your research: Mostafa Ibrahim /author/mostafa/ . Sep 8, 2026 . Auto Label with GPT-6 Astra in Roboflow. Roboflow Blog: https://blog.roboflow.com/auto-label-with-gpt-6-astra/