Which card recognition API should I use in 2026?

The short answer by use case

• You price or list on Cardmarket, or your users sell in Europe: NeoSatoshi. Every matched card comes back with its Cardmarket product ID (idProduct) and the Cardmarket trend, low, 7-day and 30-day average prices in EUR, refreshed daily.
• You need many games, sports cards or comics in one API: Ximilar or CardSight. Check their game and language lists against what your users actually have.
• You need a condition grade (corners, edges, surface, centering): Ximilar documents a separate Card Grading service. NeoSatoshi never judges condition.
• You want prices from sold listings on eBay and other marketplaces: CardSight says its prices are completed auctions, Buy It Now prices and live listings from eBay, Fanatics Collect, COMC and other public marketplaces.
• You only need a rough card name for a few photos: a general vision model can be enough, as long as you need no product ID and no price.

The options side by side (as of November 2026)

The competitor facts below come only from each company's own public pages, checked on 1 November 2026 and linked under each section. Where their pages do not say something, we write "not stated in their public docs" instead of guessing. Prices and features change, so check the linked pages before you decide.

NeoSatoshi Detection API

• Games: Pokémon, One Piece, Disney Lorcana, Yu-Gi-Oh! (TCG) and Magic: The Gathering. Pokémon is the default; the other games are chosen with one field, "tcg".
• Card languages: Pokémon in English, German, French, Italian, Spanish, Japanese, Korean and Simplified Chinese; One Piece in English and Japanese; Lorcana in English, German, French and Italian; Yu-Gi-Oh! in English, German, French, Italian, Spanish and Portuguese; Magic in English and German.
• Cards per photo: several for Pokémon, including binder pages (best up to about 20 cards per photo); one card per photo for the other four games.
• Response: name, set, number, rarity, language, Cardmarket product ID, Cardmarket prices in EUR (trend, low, avg7, avg30), TCGplayer prices in USD for Yu-Gi-Oh! and Magic, reference images, a confidence value and alternative matches. Prices are refreshed daily.
• Pricing model: one card scan per detected card, from the same balance as the NeoSatoshi app; images without a card cost nothing. Your own API key comes with any paid plan: Start-Up, the smallest paid plan, costs 19 per month (15 per month paid yearly, the same number in EUR, USD or CHF) and includes 1,000 card scans per month.
• Free testing: a public demo key, printed in the documentation, allows 10 detections per day per network without an account.
• Integration: one synchronous REST endpoint, JSON in and out, the image as base64 or as a public link (image_url), and an OpenAPI spec.

Source: Detection API documentation.

Ximilar Collectibles Recognition

• Games: their documentation says the TCG endpoint supports "cards from over 15 games, including Pokémon, Yu-Gi-Oh!, Magic: The Gathering, One Piece, and Lorcana". Separate endpoints cover sports cards, comics and sealed products.
• Card languages: with a language option it identifies the language of Latin-alphabet cards (for example English, German, Spanish); for some games it also identifies Korean, Thai, Japanese, Simplified Chinese and Traditional Chinese cards. The full list is on their taxonomy page.
• Cards per photo: by default the largest card in each image; an analyze_all option identifies every visible card in one image, charged per card.
• Response: fields such as name, set, set code, card number, rarity and year, a foil/holo tag, and for many cards links to sites such as tcgplayer.com or scryfall.com. Optional price statistics (minimum, maximum, mean and more, overall, graded or ungraded). Cardmarket product IDs are not stated in their public docs.
• Graded cards and condition: slab label reading (grade, certificate number) for Beckett, CGC, PSA, SGC, ACE, MANA and TAG, at extra credits. A separate Card Grading service grades corners, edges, surface and centering and returns a condition label.
• Pricing model: monthly API credits. Their pricing page lists a Free plan with 1,000 credits per month and Business plans from 59 EUR (69 USD) per month for 100,000 credits; identifying a TCG card costs 10 credits. A 2023 tutorial on their blog says the card identifier needs at least the Business plan, so check which plan you need today.

Sources: Ximilar recognition docs, Ximilar card grading docs, Ximilar pricing.

CardSight

• Games: their homepage lists baseball, football, basketball, hockey, MMA and soccer cards, plus Pokémon, Magic: The Gathering and One Piece. Yu-Gi-Oh! is marked "Coming Soon". Lorcana is not mentioned on their homepage.
• Card languages: Pokémon and Magic in English, French, German, Italian, Spanish, Japanese, Korean and Chinese; One Piece in English, French, Japanese, Korean and Chinese.
• Cards per photo: their homepage advertises multi-card detection in a single image.
• Response: their homepage says it returns the player, set, year and card number. Prices are completed auction results, Buy It Now prices and live listings from eBay, Fanatics Collect, COMC and other public marketplaces. TCGplayer or Cardmarket product IDs are not stated in their public docs.
• Graded cards and condition: graded and raw cards, naming PSA, BGS, SGC, CGC and TAG, and PSA population reports in beta. Condition assessment of raw cards is not mentioned.
• Pricing model: monthly API calls. Free with 750 calls per month and no credit card; Pro at 14.95 USD per month for 5,000 calls; Premium at 74.95 USD for 30,000; Ultra at 199.95 USD for 100,000; Enterprise on request.

Sources: CardSight homepage, CardSight pricing, CardSight documentation.

Building your own with open-source models

• How it usually works: a detection model finds each card in the photo, then an image embedding or text recognition (OCR) of the set code and number matches it against a reference catalog.
• The model is the smaller part. You also need a reference image for every printing in every language you support, a mapping to marketplace product IDs, a price feed you are allowed to use, and someone who adds every new set on release day.
• It makes sense when card recognition is your core product, or when you need one game in one language and control how the photos are taken.
• Cost: no fee per card, but hosting, data licensing and ongoing engineering time.

General vision models (GPT, Claude, Gemini)

• They can often name a card from a clear photo, also for games that no card API covers.
• They return no marketplace product ID, do not pick the exact printing from a catalog (set, number, language, reverse holo, edition) and have no daily price. A price they mention comes from training data or a web search, not from a price feed.
• Without structured output the answer is free text, and the same photo can give different answers on two runs.
• Cost is per token (image, prompt and answer), so it depends on the model and the image size, not on the number of cards.
• A practical pattern: let the assistant call a card API as a tool. NeoSatoshi accepts image_url, so an assistant can test it with a link and the public demo key.

How to choose: five questions

1. Which games and which card languages do your users have? Compare exact lists, not "TCG" in general.
2. Which marketplace do you price and list on? A Cardmarket workflow needs Cardmarket product IDs, a TCGplayer workflow needs TCGplayer IDs.
3. Do your photos show one card or many (binder pages, cards laid out on a table)?
4. Do you need the exact printing (reverse holo, 1st Edition, set code and rarity) or only the card name?
5. What are you charged when nothing is found? NeoSatoshi charges per detected card and nothing for an image without a card.

When not to use NeoSatoshi

• Your workflow is US-only and built on TCGplayer SKUs. NeoSatoshi keys every card on the Cardmarket product ID. TCGplayer prices in USD come only for Yu-Gi-Oh! and Magic, and there are no TCGplayer SKUs per condition.
• You need a condition grade. NeoSatoshi never judges centering, corners, edges or surface. Every card starts as Near Mint and the seller sets the condition. Slab label data (grading company, grade, cert number) is not part of the public API either.
• You need sports cards. They are not in the public API.
• You need another game, Pokémon cards in Traditional Chinese, or Magic cards in languages other than English and German.
• You need several cards per photo for One Piece, Lorcana, Yu-Gi-Oh! or Magic. Only Pokémon reads several cards in one photo.
• You need answers in well under a second, for example from a live video feed. A warm request takes about 5 seconds, a cold one 10 to 15 seconds.

How to test card recognition APIs fairly

Accuracy figures on vendor pages are measured on each vendor's own test photos, so they do not compare with each other. Collect 50 to 100 of your own real photos, with the games, languages, lighting and special printings your users have, and run every API on the same set. Score the right card, the right printing, the right language and a usable product ID. If you price by printing, count a correct name with the wrong printing as wrong.

What we measured on NeoSatoshi

• Yu-Gi-Oh!: on 315 real eBay listing photos (October 2026), the correct card in 99.3% of them.
• Magic: The Gathering: median accuracy of 97.6% in our internal evaluation (October 2026).

These are our own measurements on our own test photos. We have not measured the other APIs, so this is not a comparison.

Frequently asked questions

Is there a free card recognition API?

All three named APIs can be tried for free as of November 2026. NeoSatoshi has a public demo key (10 detections per day per network, no account). CardSight has a free tier of 750 calls per month. Ximilar lists a Free plan with 1,000 credits per month; check whether it includes card identification. For production use, NeoSatoshi needs a paid plan for your own key.

Which card recognition API returns Cardmarket product IDs?

NeoSatoshi returns the Cardmarket product ID (idProduct) for every card it can match, together with Cardmarket prices in EUR. For Ximilar and CardSight, Cardmarket IDs are not stated in their public docs as of November 2026.

Which card recognition API supports Disney Lorcana?

NeoSatoshi reads English, German, French and Italian Lorcana cards and returns foil and non-foil Cardmarket prices, see the API for Magic, Lorcana, Yu-Gi-Oh! and One Piece. Ximilar names Lorcana in its documentation. CardSight does not mention Lorcana on its homepage as of November 2026.

Can ChatGPT, Claude or Gemini identify trading cards?

They can often name a card from a clear photo. They do not return a marketplace product ID, the exact printing from a catalog or a daily price, so for listing or pricing you still need a card database. An assistant can call the NeoSatoshi API with a link and the demo key; see how to build a trading card scanner app.

Does any card recognition API grade condition?

Ximilar documents a Card Grading service for corners, edges, surface and centering. CardSight's public pages do not mention condition assessment of raw cards. NeoSatoshi does not judge condition at all.