Llama 4 Scout
Llama 4 Scout — Meta's open-weight model with MoE architecture and 10M context.
| Type | Price / 1M tokens | Notes |
|---|---|---|
| Input | $0.18 | Per 1M input tokens |
| Output | $0.59 | Per 1M output tokens |
Available free via Meta's API or self-hosted. Llama 4 Scout's mixture-of-experts design makes it efficient for long-context tasks.
Best for
- Teams with GPU infrastructure to self-host
- Ultra-long context tasks (up to 10M tokens)
- Applications needing open weights for customization
Not ideal for
- Teams without GPU infrastructure
- Tasks requiring maximum quality
Real cost scenarios
| Scenario | Est. monthly cost | Breakdown |
|---|---|---|
| Self-hosted — 2× A100 80GB (RunPod) | $584 | $0.80/hr per GPU × 730hr × 2 = $1,168 but shared across all workloads |
Estimates with typical token distributions. Use the Prompt Cost Calculator for your exact workload.
How to cut costs on Llama 4 Scout
- Self-hosting Llama 4 Scout eliminates per-token costs. Break-even vs GPT-5 mini is roughly 5-10M tokens/month depending on GPU tier.
Calculate your Llama 4 Scout costs
Other Meta models
Compare Llama 4 Scout
Frequently Asked Questions
Is Llama 4 Scout truly free?
The weights are free (open-weight). Inference has compute costs: you pay for the GPU time, not per token. Via Meta's own API, there is no charge currently. Via inference providers (Together, Fireworks, Groq), pricing varies. For high volume, self-hosting is cheapest.
What is mixture-of-experts (MoE) and why does it matter?
MoE models activate only a subset of parameters per token (Llama 4 Scout activates ~17B of its 109B total parameters). This means it runs at the cost of a 17B model while maintaining quality closer to a full 70B+ model. For inference cost and speed, this is a significant advantage.
Pricing verified 2026-08-19 from Meta official pricing. All prices in USD per 1M tokens. See methodology and price changelog.