Meta Llama API Pricing
Meta is the odd one out on this site: it does not sell a first-party paid API. Llama ships as open weights, so what you pay is set by whoever hosts it, or by your own GPU bill.
| Model | Input / 1M | Output / 1M | Cached input / 1M | Batch | Notes |
|---|---|---|---|---|---|
| Llama 4 Scout | $0.08 | $0.30 | N/A | N/A | Smaller open-weight Llama 4 tier. Hosted rates vary by provider; self-hosting shifts the cost to GPU time instead. |
| Llama 4 Maverick | $0.20 | $0.60 | N/A | N/A | Meta does not run a first-party paid API. These are representative rates from hosted providers; the exact figure varies by host, so check yours. |
Meta's cheapest tracked model is Llama 4 Scout at $0.08 in / $0.30 out, and its most expensive is Llama 4 Maverick at $0.20 in / $0.60 out, a 2.2x spread across the lineup. Llama 4 Maverick ranks #14 of 60 models tracked across 17 providers.
There is no single Llama price
The rates on this page are representative figures from hosted providers, not a price list Meta publishes. The same Llama 4 Maverick weights can cost meaningfully different amounts depending on which host you route to, what quantisation they run, what context length they allow, and how aggressively they are competing on price that month. Treat these numbers as the middle of a range rather than a quote.
This is a real advantage and a real hassle at once. The advantage is that hosting is a commodity: if one provider raises prices you can move, and the model itself does not change. The hassle is that comparing Llama against a closed model is comparing a market to a price tag.
Hosted or self-hosted
For anything below roughly continuous load, hosted Llama almost always wins on cost, because a rented GPU bills whether or not you are sending it traffic. Self-hosting starts to make sense at sustained high utilisation, or when data residency or latency requirements make a third-party host a non-starter.
The break-even is arithmetic, not opinion: put your monthly token volume into the GPU rental cost calculator and compare against the hosted rate. Remember to include the engineering time to run it, which is the line item that most often makes self-hosting the more expensive option in practice.
Where Llama sits on price
Llama 4 Scout and Maverick both land in the cheap end of the table, comparable to Gemini Flash-Lite and the smaller Grok and GPT tiers. They are not competing with frontier models on capability, and they are not priced as if they were.
Every Meta model on this site
| Model | Blended / 1M | Rank of 60 | Full pricing |
|---|---|---|---|
| Llama 4 Scout | $0.135 | #6 | Llama 4 Scout pricing page |
| Llama 4 Maverick | $0.300 | #14 | Llama 4 Maverick pricing page |
Meta pricing FAQ
Does Meta charge for the Llama API?
Meta does not operate a paid first-party Llama API. It publishes the weights, and third-party hosts sell inference. Any Llama price you see, including the ones on this page, is a host's price rather than Meta's.
Is it cheaper to self-host Llama than to use a hosted API?
Only at sustained high utilisation. A rented GPU bills by the hour whether or not you send it traffic, while a hosted API bills per token. Work out the crossover with the GPU rental cost calculator, and include your own engineering time, which is usually what tips the balance.
Why do different providers charge different prices for the same Llama model?
Because they are selling infrastructure, not the model. Quantisation, batching strategy, hardware, context limits, and how hard they are competing for your traffic all move the number. The weights are identical; the service around them is not.