How Much Does It Cost to AI-Parse a Menu?

When people hear that Travel Eat runs every menu through a large AI model — extracting each dish with its price, ingredients, allergens, dietary type, and nutrition estimate — the first question is usually some version of "doesn't that cost a fortune?"

The total depends on the selected Actor, current platform event rates, and any model usage billed to your own provider account. Here is how to separate those costs before estimating a run.

What "parsing a menu" means here

To be clear about what you're pricing: this isn't OCR. For every dish, the pipeline extracts:

• Name (translated) and original name as printed
• Category and price, with currency
• Atomic ingredients — pesto gets decomposed into basil, pine nuts, parmesan, olive oil — each with EU-14 allergen flags and confidence levels
• Dietary type (vegan / vegetarian / pescetarian) with an explanation
• Estimated nutrition per serving
• A one-line origin story, because food is culture

That's a lot of structured output per dish, and output tokens are the expensive kind. Keep that in mind — this is the maximalist version of menu parsing, and it still costs pennies.

The math

Our pipeline runs on Gemini 2.5 Flash by default. Take a typical menu: six photos, about 35 dishes.

Input: six menu photos plus our parsing prompt come to roughly 10,000 tokens. At Google's published Flash input rate (about $0.30 per million tokens as we write this), that's $0.003.

Output: 35 dishes at a few hundred tokens each is roughly 9,000 tokens of structured JSON. At about $2.50 per million output tokens, that's $0.022.

So the AI bill for the entire menu is around 2–3 cents — well under a tenth of a cent per dish. Simpler menus, or the Flash-Lite model (which is what our photo matcher uses), come in several times cheaper still. These model rates are an illustrative July 2026 snapshot. Check Google's current pricing before estimating an actual run.

Why we made it bring-your-own-key

Our public Apify Actors offer managed and BYOK billing so you can choose how model usage is paid. Compare current platform event rates and provider usage charges before selecting a variant.

The BYOK variant lets you bring your own Gemini key (free to create). Model usage is billed to your Google account under its current pricing and quotas; any free-tier allowance depends on that account. The AI Restaurant Menu Parser also charges platform events. Check its Pricing tab for the current event rates.

For the illustrative 35-dish menu above, estimate the total by adding the Actor's current event charges to your provider's model-usage charges. With BYOK, your provider account controls the model quota and reports that part of the cost.

The whole toolkit

The parser is one of three actors we've published from Travel Eat's production pipeline:

• AI Restaurant Menu Parser — menu photo URLs in, structured dishes out. Check current event rates before estimating a run.
• Google Maps Menu Scraper — don't have photos? Start from a Google Maps place URL; it finds the menu photos and parses them in one run.
• Restaurant Dish Photo Matcher — attach the place's real food photos to each parsed dish, with pricing shown on its Store page.

If you'd rather skip key setup, each Actor has a managed twin with API keys included: Menu Parser, Menu Scraper, and Photo Matcher. Compare each current Pricing tab with the BYOK variant and your provider's usage charges; a managed variant does not necessarily cost more.

Structured menu data is available through these APIs. Choose a billing variant after comparing current platform rates, provider usage charges, and the size of your menu.

Travel Eat scans, translates, and allergen-checks restaurant menus in 36 languages — powered by exactly this pipeline.