Menu OCR
Scan a menu photo or PDF to JSON: every dish with its section, description, price and options, ready for a delivery app or POS.
3 free pages a day, no sign-up.
Batch process or save historyWhat you get back
The sample menu above, read with document_type=menu. The first 8 of its 17 dishes are shown.
{
"document_type": "menu",
"fields": {
"restaurant": "Harbor Lane Kitchen",
"currency": "USD",
"sections": [
"Starters",
"Salads",
"Desserts",
"Wood-fired pizza",
"Mains",
"Drinks"
],
"items": [
{
"section": "Starters",
"name": "Crispy Calamari",
"description": "Lemon aioli, charred lemon",
"price": 12.5,
"currency": "USD",
"options": []
},
{
"section": "Starters",
"name": "Burrata & Heirloom Tomato",
"description": "Basil oil, sourdough crisps",
"price": 14,
"currency": "USD",
"options": []
},
{
"section": "Starters",
"name": "Soup of the Day",
"description": "Ask your server",
"price": 6,
"currency": "USD",
"options": [
"Cup 6",
"Bowl 9"
]
},
{
"section": "Salads",
"name": "Little Gem Caesar",
"description": "Anchovy dressing, croutons",
"price": 11,
"currency": "USD",
"options": [
"Add chicken +5",
"Add shrimp +7"
]
},
{
"section": "Wood-fired pizza",
"name": "Margherita",
"description": "San Marzano tomato, fior di latte, basil",
"price": 16,
"currency": "USD",
"options": [
"Add burrata +5"
]
},
{
"section": "Wood-fired pizza",
"name": "Wild Mushroom",
"description": "Fontina, thyme, truffle oil (v)",
"price": 18,
"currency": "USD",
"options": [
"Gluten-free crust +3"
]
},
{
"section": "Mains",
"name": "Half Roast Chicken",
"description": "Lemon, garlic, herbs",
"price": 26,
"currency": "USD",
"options": [
"Choice of fries or salad"
]
},
{
"section": "Drinks",
"name": "Glass of Wine",
"description": "Red, white or rosé",
"price": 12,
"currency": "USD",
"options": [
"Bottle 44"
]
}
]
}
}| Section | Dish | Price | Options |
|---|---|---|---|
| Starters | Crispy Calamari Lemon aioli, charred lemon | 12.50 | |
| Starters | Burrata & Heirloom Tomato Basil oil, sourdough crisps | 14.00 | |
| Starters | Soup of the Day Ask your server | 6.00 | Cup 6, Bowl 9 |
| Salads | Little Gem Caesar Anchovy dressing, croutons | 11.00 | Add chicken +5, Add shrimp +7 |
| Wood-fired pizza | Margherita San Marzano tomato, fior di latte, basil | 16.00 | Add burrata +5 |
| Wood-fired pizza | Wild Mushroom Fontina, thyme, truffle oil (v) | 18.00 | Gluten-free crust +3 |
| Mains | Half Roast Chicken Lemon, garlic, herbs | 26.00 | Choice of fries or salad |
| Drinks | Glass of Wine Red, white or rosé | 12.00 | Bottle 44 |
Menu OCR benchmark
We ran 6 test menus through 5 models on 2026-11-19, then extracted dishes with document_type=menu. Every model found every dish and read every price exactly, so on clean menus the choice comes down to speed and price: anthropic/claude had the fastest OCR, openai/gpt the lowest price.
| Model | Dishes found | Prices exact | OCR time (median) | Total with fields | Credits per page |
|---|---|---|---|---|---|
| google/gemini-flash | 70 / 70 | 70 / 70 | 34.0 s | 51.8 s | 26 |
| mistral/mistral-ocr | 70 / 70 | 70 / 70 | 6.4 s | 20.3 s | 36 |
| openai/gpt | 70 / 70 | 70 / 70 | 4.2 s | 16.2 s | 22 |
| anthropic/claude | 70 / 70 | 70 / 70 | 2.8 s | 14.8 s | 36 |
| azure/document-intelligence | 70 / 70 | 70 / 70 | 8.2 s | 23.4 s | 71 |
Test menus: single column, printed; two columns, printed (the page sample); phone photo of the two-column menu (angle, shadow, blur); chalkboard, handwriting-style fonts; japanese (non-latin); pdf menu. All are made up and generated, so they are cleaner than most real photos; read this as a floor, not a ranking. A dish counts as found when its name matches, and a price as exact when it equals the printed number. Credits include the 11 for fields; times include upstream queueing on the day. For harder real-world menus, see the academic benchmark Evaluating Menu OCR and Translation (arXiv 2504.13945) and DoorDash's engineering post on transcribing restaurant menu photos with LLMs.
Which models read menus best
All five read our test menus without a wrong dish or price (see the benchmark above), so pick by speed, price and the scripts you need. Try two on your own menu; the model is one setting.
| Model | Handwriting | Languages | Per page |
|---|---|---|---|
Gemini Flashgoogle/gemini-flashEvery dish and price right in our test. Our default for menus, though the slowest of the five. | Yes | 100+ languages | 15 credits |
Mistral OCRmistral/mistral-ocrClean Markdown with tables for printed menus and PDFs. Handwriting is not in its capabilities, so skip it for chalkboards. | No | 100+ languages | 25 credits |
GPTopenai/gptGood on messy photos. The lowest price per page of the five. | Yes | 50+ languages | 11 credits |
Claudeanthropic/claudeCareful on long, dense menus such as wine lists. The most expensive of the three vision models. | Yes | 50+ languages | 25 credits |
Azure Document Intelligenceazure/document-intelligence300+ languages and handwriting, for menus in scripts the vision models read less often. | Yes | 300+ languages | 60 credits |
Pricing per page
- Reading the menu15 credits per page with Gemini Flash, other models as listed above
- Dishes, prices and sections (document_type=menu)11 credits per file
- A one-page menu with Gemini Flash26 credits, about $0.026
1,000 credits cost $1. Failed requests are not charged, and sending the same file again reuses the OCR result for free. See plans
Do it with the API
Converting more than a few files? The same conversion is one request. Pick any model, or let auto choose.
curl https://api.ocrrouter.ai/v1/ocr \
-H "Authorization: Bearer $OCRROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "google/gemini-flash",
"file_url": "https://example.com/menu.jpg",
"format": "markdown",
"document_type": "menu"
}'