Depth Anything 3 small API pricing: the dated answer, seller by seller, unit by unit

There is no official Depth Anything 3 API to price. The team behind the models, at ByteDance Seed, ships weights and code. It ships nothing hosted. The small tier's page on Hugging Face lists zero inference providers as of 26 October 2026. What you can buy is a scatter of units. Community rows on Replicate bill by GPU second. One V2 endpoint bills per image. A video model bills by length. A browser workspace bills credits. This page maps every seller with its unit, its date, and its license. Then it works a per-image budget on this platform's da3-small lane, which has served since 6 October 2026.

Last verified: 2026-10-26

What you are actually trying to price

Depth Anything 3 is a depth and camera pose model family from a ByteDance Seed team. It shipped in November 2025. The paper reports it beating the prior multi-view state of the art by 44.3 percent on camera pose accuracy, on average. Training used public academic data only. The small tier, DA3-SMALL on Hugging Face, is the one this page prices. It holds 34.3 million parameters, a 137 MB weights file, and the Apache 2.0 license. One call sends one or more images. It returns a relative depth map for each view, plus confidence and camera parameters. The full record for the model lives on its own page here: da3-small.

The unit decides everything. Depth work is counted in images. One image in, one depth map out. A price for this model is a price per image, the way a chat price is per token and a robot price is per call. When a seller bills something else, seconds or credits or video length, your first job is conversion. The sections below do that with dated numbers.

One naming note. Search results mix the releases freely. The publisher writes DA3. Searchers type "depth anything v3 api" and "depth anything 3 api cost". Both spellings mean the 2025 release. The suggest box also pairs the family with the plainer string "depth map api". This page answers to all of them.

The official answer is no API, and the searches show it

The honest headline for "depth anything api pricing" is that the makers sell nothing. The official repository links code, weights, and docs. There is no product page, no account, no checkout. The repo does carry a page called an API, in its docs folder. That page documents a Python class you import on your own machine. It is not a service you call over the network. Some searchers get stuck here.

The demand side is visible in the searches themselves. Google's suggest box, checked 26 October 2026, finishes "depth anything api" with the string "depth anything no api found". It finishes "depth anything v2 api" with "depth anything v2 no api found" and "depth anything v2 error no api found". People are not only asking for a price. Enough of them failed to find any API that the failure itself became a suggested search. The machine surface agrees. The Hugging Face provider map for DA3-SMALL returns an empty list, read the same day. Zero hosted providers on the platform most people check first.

The ecosystem's answer today is to build your own. Users publish their own endpoint repos. Wrapper sites resell runs. Guides point at the community rows. That is the market this page prices.

The seller census, checked 26 October 2026

Every row below was read on its own page today. Rates move, so every price carries its unit and its date. The empty rows carry their checks.

Three notes on the table. First, the DA3 community rows are real, but they price hardware time, not images. A GPU second price cannot become a per-image price until you know how long one image takes on that card. Sellers' median run figures are their own estimate of that, at their default settings. Second, the rows reprice. An October 2026 read of the cjwbw row showed about 0.053 per run on L40S hardware. Today that row reads differently, at a lower median on the same hardware class. Treat every rate as a snapshot with a date, ours included. Third, the wrapper and workspace rows serve the older releases. Nobody in the table sells the DA3 small tier per image, except the platform whose fence sits below.

The empty rows are checked, not assumed. Two new robot platforms appeared in late 2026, abstract.inc and Robotic Inference. Both were read today. The first publishes a robot rate card with no depth row. The second carries no depth mention in its render. The absence is dated, and any row could change.

The license decides what you may ship

Price is half the question for a depth model. The other half is what you may ship in a product. This family splits its licenses by tier, and most pricing pages never mention it. The V2 repository states its split in one sentence, read 26 October 2026. The small model is Apache 2.0. The base, large, and giant models carry the non-commercial CC BY-NC 4.0 license. The DA3 side, read the same day from the repository's model table:

The trap is quiet and cheap to avoid. Community rows sell runs of whichever tier the row builder picked. A price on a row does not fix a license on weights. If you ship a product, you need an Apache tier, whatever you pay per run. One live disagreement between the official surfaces, found on 26 October 2026: the refreshed DA3 large 1.1 card on Hugging Face says Apache 2.0. The repository's own table row says CC BY-NC 4.0. That is why the rule is to read the license block on the exact tier page you download from, on the day you download it. The same check across the other robot families is walked in the GR00T license guide.

What a per-image price looks like

The query "depth estimation api price per image" presumes a market that bills per image. Mostly, that market does not exist yet. The sellers bill seconds, credits, or video length, as the census shows. Depth work is cheap compute either way. One pass of the small tier takes milliseconds of GPU time. So a per-image depth price follows the band around image APIs, not the cost of the math. The October 2026 band among the major image APIs, read from the publishers' price pages for our rate work: Imagen 4 standard at 0.04 per image, Gemini image output about 0.045, Flux 1.1 pro at 0.04, GPT-Image-2 medium about 0.053. Those products generate images. That costs far more compute than measuring depth. They cluster tightly anyway.

Against that band, here are today's reads for this family. The one per-image seller in the census, ModelRunner, prices V2 at 0.01 per image. That is one order below the band, for the older release. The platform rate in the fence below prices the DA3 small tier per image, inside the band. Anyone quoting you a depth number should also name the unit, the party, the date, and the tier license. A number without those four is not a quote.

The platform rate and the budget math

This platform serves the DA3 small tier as a metered lane under the da3-small model id, on its vision route. The row is on the public list as of 26 October 2026. It has served since 6 October 2026. Metering is per image, counted from the image parts in your request. The route takes multipart form data, one or more image parts. It returns the model's output plus a usage block. The usage block carries the billed image count and the cost of the call. A request refused before it is served bills nothing. Every field and error code is in the run-vision reference. The vision route page walks the contract end to end.

Budgeting the lane is math on two named inputs:

monthly spend  =  N x P
worst case     =  the prepaid wallet balance W

N: images per month, from your workload
P: price per image, read off the live table above for the da3-small row
W: wallet balance, the hard stop on spend

The numeric form, at the fenced rate. A labeling batch of 10,000 images a month bills 500. A loop that sends one image per second, four hours a day, 22 days a month, sends 316,800 images and bills 15,840. Per-image billing on a robot loop is a real budget line. Frame rate is a cost lever, not a detail. A wallet holding 10 buys 200 images and stops there. One sensitivity line: the spend is linear in images and nothing else. Send half the frames, pay half. A larger image costs the same as a smaller one.

The wallet stop is the point of the shape. An unattended loop cannot run up a credit line. A refused request bills nothing. The full argument for prepaid floors under robot and agent workloads, with runaway cases, is the API bill shock guide. For how depth pairs with masks in a robot perception stack, see the robot vision guide.

What to do now

1. Settle the license before the price. Pick your tier from the license table above. Then read the license block on that tier's own page, on the day you download. That fixes what you may ship, before any rate matters.
2. Convert every quote to one unit. For the GPU second rows, cost per image equals the per-second rate times the seconds one image takes on that card. The seller's median run price is that product, worked out at their settings. It moves with input size.
3. Write the client against a published contract. The vision route's fields, limits, and error codes are fixed and written down today. Your client can be finished and tested before any spend begins. The models index is the live catalog.
4. Budget from the live table. Listed rates render on the pricing page. The signup flow opens the wallet and the keys. The prepaid balance is the spend stop.

Frequently asked questions

Is there a Depth Anything 3 API?

No official one. The ByteDance Seed team ships weights, code, and a local Python interface. It ships nothing hosted. Checked 26 October 2026: the Hugging Face provider map for the small tier is empty, and the makers' pages sell nothing. What exists are community rows on Replicate, three DA3 tiers, billed by GPU second. One V2 endpoint bills per image on ModelRunner. Leap wraps the video model. A credits workspace sells browser runs. This platform bills per image for the small tier. Google's suggest box still finishes "depth anything api" with "no api found". This section is the dated answer to that search.

What does Depth Anything 3 cost per image?

No first-party price exists. In the census read 26 October 2026, the DA3 rows bill 0.000225 to 0.000975 per GPU second, with median runs from about 0.00098 to about 0.020. Your per-image cost depends on the time one image takes on the seller's card. The one clean per-image read in the family is the older V2 release at 0.01 per image on ModelRunner. This platform's rate for the DA3 small tier renders in the table above and on the pricing page. The worked math above turns it into a monthly number.

Why do Replicate prices show per second, not per image?

Because those rows bill hardware time. Replicate meters the seconds your run occupies the card. The page shows a per-second rate plus a median run price as the budgeting aid. To compare it with any per-image quote, multiply the rate by the seconds one image takes on that hardware. Check the input size too. Time per image scales with image size, and the seller's median is only their default.

Is Depth Anything V2 free for commercial use?

Only the small tier. The V2 repository states it in one sentence. The small model is Apache 2.0. The base, large, and giant models are non-commercial under CC BY-NC 4.0. This bites on the cheap rows. Their default download is often the large tier. If you ship V2 in a product, name the small tier, or move to the Apache tiers of DA3.

Can I run DA3 small myself instead of paying per image?

Yes, and that path is always open. The small tier holds 34.3 million parameters and a 137 MB weights file, under Apache 2.0. The card's own example code runs on CPU when no GPU is present. Self-hosting trades the per-image bill for your own serving work: the process, the uptime, the fixes. The card-rate side of that trade is worked in the local versus hosted guide. The budget math above is the number self-hosting has to beat.

The short version: no official API. Five community rows bill GPU seconds, one endpoint bills V2 per image, one wrapper bills video length, and one workspace bills credits. The license split rules most community tiers out for products. This platform's DA3 small lane bills per image. The live rate is on the pricing page. You can sign up and create a key without a call from us.

Last verified: 2026-10-26