Llama 4

Llama 4 is Meta's open-weight model family, released in Scout and Maverick variants on a Mixture-of-Experts architecture with native text-and-image multimodal understanding. The most important thing to understand before comparing its "price" to a closed model like GPT or Claude is that Llama 4 doesn't have a first-party price at all: Meta doesn't sell inference for it, and the model weights are downloadable for free from Hugging Face under the Llama 4 Community License.

That license permits commercial use below a 700-million-monthly-active-user threshold, requires "Built with Llama" attribution, and prohibits using the model's outputs to train a competing model or building a product that directly competes with Meta's own core businesses (social networking, messaging, AR/VR). It's also worth noting the license currently restricts direct access to Llama 4's multimodal capabilities for individuals or companies based in the EU, though that restriction doesn't extend to end users of products built on top of it elsewhere.

Because Meta doesn't host inference, running Llama 4 in practice means one of two paths: self-hosting the downloaded weights on owned or rented infrastructure, or using a third-party inference provider (options include Groq, Together AI, Fireworks, Replicate, DeepInfra, Novita, and OpenRouter, among others) that hosts the model and charges per token. Pricing across those providers varies meaningfully by speed and provider margin rather than by anything Meta controls, so "how much does Llama 4 cost" only has an answer once a specific provider and workload are chosen. Cheaper providers on cost-per-token can run several times less than what a comparable closed frontier model charges per token, but the tradeoff is usually inference speed, uptime guarantees, or the operational burden of comparing providers directly before settling on one.

The Scout variant's headline feature is a very large context window (reported at up to 10 million tokens), aimed at workloads like analyzing long documents, large codebases, or retrieval-heavy pipelines in a single pass rather than chunking input across multiple calls. Maverick trades some of that context length for stronger general benchmark performance, at the cost of needing more infrastructure to run well. In practice, effective usable context tends to degrade at the extreme end of the advertised window, so the 10M figure is closer to a ceiling than a guaranteed working range for every task, and workloads that actually need the full window should be tested rather than assumed to work at face value.

For someone evaluating whether an open-weight model is the right choice at all, the honest tradeoff is control and cost-per-token flexibility (self-hosting, provider shopping, no vendor lock-in) against the convenience of a single flat subscription and a model provider handling infrastructure, uptime, and support directly. Teams with in-house infrastructure expertise or high enough token volume to make provider-shopping worthwhile tend to get the most out of an open-weight model like Llama 4; teams that just want a predictable monthly bill and no infrastructure decisions to manage are usually better served by a closed model's consumer subscription or managed API instead, where that operational overhead is already handled by the vendor.