The homepage says Roboflow turns images, videos, and streams into intelligence for companies that run in the physical world. Plans start at $0. A Free Tier lists 10 credits included a month enough to train ~30 models or run 80,000 inferences, and Core is the paid monthly step. The pricing description says free self-hosted inference and a model-weight download sit on every plan. Those credit lines are the vendor’s estimate of what a credit can cover.
Official website · roboflow.com
Use Roboflow when you have images or video to label, a vision model to train, and a place to run it: Roboflow’s cloud, or your own NVIDIA Jetson, Raspberry Pi, or GPU. The Free Tier is $0. Core is $39 per month, billed monthly beyond the Free Tier, and the first pack under that price is 20 credits in total at $3.90/credit. Skip it if you wanted a general chat product, or if commercial self-hosting of a model such as RF-DETR 2XL is required on Core: that extra license is marked not included there. The “enough to train” sentences are not a training log from this profile.
The homepage’s platform section says workflows combine models, orchestrate actions, update databases, and plug into industrial systems. The drawn example lists an object-detection model, motion detection, a ByteTrack tracker, a line counter, an email notification, and Roboflow Vision Events. Treat that stack as the page’s illustration. It is not a camera feed this profile ran. Around that diagram, the homepage names jobs such as catching defects, predicting maintenance, improving warehouse safety, gaining logistics visibility, inspecting infrastructure, guiding machines, verifying prescriptions, and checking food specifications. It also states that over half the Fortune 100 and 80,000 organizations go from idea to production. It says over 16,000 organizations build with Roboflow, and it includes quotes from BNSF and Pella. Those figures and quotes are the company’s, not an independent count.
Official documentationThe annotate section says you can annotate thousands of images in minutes using frontier models and custom filters. In the docs, datasets are where you upload, manage, annotate, and analyze images, and a models blurb says you train state of the art models with a few clicks. Deployments on that page run models in the cloud or on your own hardware. A later quickstart says to run models on cloud GPUs, with no setup, to run a model locally with Docker, and to detect, track, and count objects in a video. Pricing says all plans include model inference of over 50 supported models and model training of dozens of architectures, plus active learning that iterates with every inference, workflows, vision events, and model-weight download. The same block says the workspace and models are private, and that you may optionally share to Roboflow Universe. The docs tell you to find your API key and to use a model API endpoint, and they mention a CLI and a Python SDK.
Official documentationServerless model inference is billed per image with no infrastructure to manage, on what the pricing page calls autoscaling GPUs. The rows are credits, not a dollar fee per image. RF-DETR Medium is 0.0001875 credits and 5,333 images per credit. RF-DETR Nano, YOLO26 Nano, and YOLOv8 Nano are 0.0001250 credits and 8,000 images per credit. SAM 3 is 0.0005000 credits and 2,000 images per credit. Qwen3-VL is listed as GPU time of 500 seconds, and the table points to 30+ more detection, VLM, and other models. Self-hosted inference is a separate row marked included: Roboflow Inference on your own NVIDIA Jetson, Raspberry Pi, or GPU. The pricing description repeats that this self-hosted inference and model-weight download are free on every plan. The comparison table calls Inference an open source deployment product and says hosted models come with a commercial license. Some models, like RF-DETR 2XL, still need an additional license for commercial self-hosting, and the Core cell for that license says not included while the Enterprise cell says included.
Official documentationThe homepage says to connect your coding agent, like Claude Code, Codex, or Cursor, so it can make deployable vision applications. It says to build with industry-standard open source and lists the supervision, inference, rf-detr, notebooks, and trackers repositories, each with a star count shown beside the name. Hand setup tells you to fetch the agents document and paste it into the agent. The server address on that page is https://mcp.roboflow.com/mcp. Separate notes cover Claude, Cursor, ChatGPT, Codex, and Gemini CLI, and the ChatGPT note says developer mode may be required and that availability depends on the account and workspace. The pricing page says one command installs the MCP and skills and that you use the same Roboflow plans you already have. Enterprise, beside that, is custom pricing: priority access to faster GPUs, volume pricing, an uptime SLA, a US or EU data sovereignty guarantee, email, ticket, and chat support, and workspaces that are always private. Labeling services on the add-on list start at $0.10/bounding box, $0.20/polygon, and $0.05/classification/keypoint annotation. HIPAA compliance and BAA is one line under custom contracting, not a document reproduced here.
Official documentationIs to open the Free Tier at $0 and label a narrow set, and train once while the card still says 10 credits included a month enough to train ~30 models or run 80,000 inferences. Read that as the page’s ceiling estimate, then compare it with the per-image credit rate of the architecture you pick. This suggestion is the pricing card’s sequence. If the job is larger than that estimate, the same page is where Core’s pack begins.
A suggested deployment choice is serverless when you want autoscaling GPUs billed per image, and self-hosted Inference when the stream should stay on a Jetson, a Raspberry Pi, or your GPU. The plan table marks that self-hosted runtime included, which is a different line from the model credit rows. On the same comparison, dedicated deployments, batch processing, and the serverless hosted API are included for Core and for Enterprise. Vision events keep a 14-day lookback on the Core column and a custom lookback on Enterprise. Export of those events is marked not included on Core. Use those cells as the boundary of this suggestion. They are not measurements from a live site.
A suggested developer path is to keep images in the Roboflow workspace and connect an agent named on the homepage, including Cursor. The install copy also names Claude, ChatGPT, Codex, and Gemini CLI. ChatGPT’s note says you may need developer mode, and that the option depends on account and workspace settings. The pricing page ties that install to the plans you already pay for. It does not say agent actions are exempt from credits. Check the Free Tier allotment or the Core pack before a training run. This is a suggested install order from the site, not an agent session opened here.
The homepage says to try the free tier and that the docs are open. Plans start at $0 and the pricing description says to start free with 10 credits a month.
Follow the docs path to upload, manage, annotate, and analyze images. The homepage’s annotate block is the one that mentions frontier models and custom filters.
Use the documented training path, then a model API endpoint on cloud GPUs or a run on your own hardware. Self-hosted Inference is the included row for a Jetson, Raspberry Pi, or GPU.
Stay with the Free Tier’s monthly credits, take Core when you want the $39 pack shown on the page, or contact sales for Enterprise custom pricing. Labeling-service rates apply only if you buy that add-on.
The pricing page shows Free Tier at $0 with 10 credits included a month, and Core at $39 per month, billed monthly. Under Core, the first pack line is +10 credits a month, 20 credits in total, at $3.90/credit, followed by the sentence that this is enough to train ~60 models or run 160,000 inferences. Further monthly packs are shown between that $3.90 line and the ~60-model sentence. Prepaid credits start at $4 and expire after one year. On-demand credits are $6 and require a linked credit card. Enterprise says custom pricing, so no Enterprise dollar amount is published. The homepage also says plans start at $0.
Trying labeling, training, and included self-hosting.
$0.
10 credits included a month.
A monthly credit pack beyond the free allotment.
$39.
Per month, billed monthly.
Production deployments that need the listed add-ons.
Not verified.
Custom Pricing.
The “enough to train” lines are the pricing card’s wording, not a measured result. Serverless rows bill credits per image. The extra commercial license for self-hosting models such as RF-DETR 2XL is not included on Core and is included on Enterprise. Workflow version restore and vision-events export are not included on Core. Model monitoring is not included on Core and is optional on Enterprise. Core’s vision-events lookback is 14 days. Labeling prices are starting rates for a service add-on. The pages do not state tax, a refund rule, or a numeric cap for hosted API requests per minute. They say the Core limit is standard and the Enterprise limit is custom.
You want a general model host instead of this vision dataset flow.
Roboflow’s public pages are about labeling images, training vision models, a Free Tier at $0 and deploying to cloud GPUs or your own Jetson, Pi, or GPU. Choose Hugging Face when that other listing is the product you mean to evaluate.
You want to run a published model without this labeling workflow.
Replicate is a different developer-tool listing. These Roboflow pages specify dataset tools, Core at $39 per month, and per-image credit rates for named vision models. Use Replicate when you are choosing that listing’s model-running job rather than this dataset-to-deploy path.
The work stays in the editor, not in a vision workspace.
Cursor is its own developer tool. Roboflow’s homepage names Cursor as one coding agent you can connect so it can build vision applications. Choose Cursor when you are evaluating the editor, not Roboflow’s credits, training, and deploy targets.
Yes. The Free Tier is $0 and the homepage says plans start at $0. The card lists 10 credits included a month and calls that enough to train ~30 models or run 80,000 inferences. The pricing description says to start free with 10 credits a month, and that free self-hosted inference and model-weight download are on every plan. You can upgrade to Core at $39/mo or to Enterprise. A linked card is mentioned for on-demand credits, not on the Free Tier button.
SourceCore is $39 per month, billed monthly beyond the Free Tier. The first pack line is +10 credits a month, 20 credits in total, at $3.90/credit, with a page estimate of ~60 models or 160,000 inferences. All plans include inference for over 50 models, training for dozens of architectures, the Roboflow Agent, on-device deploy, workflows, vision events, active learning, and model-weight download. Workspaces and models are private, with an option to share to Roboflow Universe.
SourceServerless inference is billed per image with no infrastructure to manage. Published examples include RF-DETR Nano at 0.0001250 credits and 8,000 images per credit, and SAM 3 at 0.0005000 credits and 2,000 images per credit. Qwen3-VL is GPU time of 500 seconds. Self-hosted Inference on your own hardware is marked included.
SourceEnterprise shows custom pricing and a contact-sales path, so that price was not verified here. Listed extras include faster GPUs, volume pricing, an uptime SLA, a US or EU data sovereignty guarantee, and email, ticket, and chat support. Workspaces and models are always private. Labeling services start at $0.10 per bounding box, $0.20 per polygon, and $0.05 per classification or keypoint annotation. HIPAA compliance and BAA appears under custom contracting.
SourceThe comparison table calls Inference an open source deployment product, and the homepage says to build with industry-standard open source. That does not say every weight file is open source. Some models, like RF-DETR 2XL, need an additional license for commercial self-hosting, which the Core column marks not included. Hosted models are said to come with a commercial license. The docs say to use a model API endpoint and to find your API key. API rate limits are standard on Core and custom on Enterprise, described as requests per minute. A number for that cap is not shown.
SourceVision-events lookback is 14 days on Core and custom on Enterprise. Export of those events, and restoring previous workflow versions, are marked not included on Core. Model monitoring is not included on Core and optional on Enterprise. Prepaid credits start at $4 and expire after one year. On-demand credits are $6 and need a linked card. The pages do not say whether unused monthly credits roll over, and they do not show tax or a refund rule. Customer quotes remain vendor statements.
SourceOct 6, 2026 · NavHunt editorial
NavHunt did not run Roboflow, did not create a workspace, and did not upload, train, or call a model. Credit rows, pack prices, and the “enough to train” lines come from the homepage, pricing page, and docs reviewed on 6 October 2026. Accuracy and real credit burn were not measured. This description is based on Roboflow's official materials reviewed on 2026-10-06, including Roboflow: AI for the physical world, Roboflow Pricing: Free, Core & Enterprise Plans, Roboflow Documentation | Roboflow Docs.
The Free Tier includes a monthly credit allotment. Core’s first pack is priced on the same page.