Because Llama is open-weight, the "model family" is really a set of model files at different sizes. Bigger models are generally more capable and cost more per token to serve. Smaller models are cheaper and faster, and are good enough for a surprising amount of everyday work.
The chart below is live. It reads hosted-API rates from our catalogue, which are sourced from OpenRouter, and it updates whenever our pricing sync runs. Use it instead of any number you remember from a blog post.
The lineup
| Model | Context window | Best for | Notes |
|---|---|---|---|
| Llama 4 Scout | 1,310,720 tokens | Long documents, cheap general use | Lowest-priced Llama 4 model; very large context |
| Llama 4 Maverick | 1,048,576 tokens | Harder tasks, better answers | Roughly double Scout's per-token price |
| Llama 3.3 70B Instruct | 131,072 tokens | Proven workhorse for chat and writing | Same input rate as Scout on OpenRouter at the time of writing |
| Llama 3.1 8B Instruct | 131,072 tokens | Classification, extraction, bulk jobs | Cheapest model in the family; also easy to self-host |
| Muse Spark 1.3 | 1,048,576 tokens | Meta's premium hosted model | Priced like a frontier model; check whether you need it |
| Muse Glimmer 30B | 131,072 tokens | Mid-priced hosted option | Sits between Llama and Muse Spark on price |
Llama vs Muse
Our catalogue lists two groups under Meta. The Llama models are the open-weight family you can download and run anywhere. The Muse Spark and Muse Glimmer models are listed as hosted-API products at much higher rates. At the time of writing Muse Spark 1.3 costs $1.25 per million input tokens against $0.20 for Llama 4 Maverick, so treat Muse as a separate decision and only pay for it if a cheaper Llama model fails on your task.
Context windows
The context window is the maximum number of tokens the model can hold at once: your prompt, any documents you paste, and its reply. Llama 4 Scout and Maverick list windows above one million tokens in our catalogue. Llama 3.3 and 3.1 list 131,072 tokens, which is still around 100,000 words.
Two cautions. First, the window you actually get depends on the host: some serve a shorter window than the model supports. Second, a big window is a cost trap. Every token you send is billed, so pasting a 300-page PDF into every message can cost more than the answer is worth. Lesson 8 covers how to avoid that.
Knowledge check
You need to classify 50,000 short support tickets as cheaply as possible. Which model should you try first?