Llama 4 Scout vs Maverick: API pricing, self-hosting costs, and which one to use

Why both models have the same active parameter count

Scout is 109 billion parameters total. Maverick is 400 billion. But inference cost is driven by active parameters, and both activate exactly 17B per token. That gap is the MoE architecture: each token gets routed to one expert out of the pool, and the rest sit idle.

Each token gets routed to 1 of 16 experts in Scout, or 1 of 128 in Maverick, plus a shared expert in both cases. The larger pool gives Maverick more specialized knowledge without changing per-token compute. Meta also trained Maverick through codistillation from Behemoth, their 2-trillion-parameter teacher model. That's the most likely explanation for why Maverick outperforms models with comparable active parameter counts on hard reasoning tasks.

Scout's design trades expert depth for context. The iRoPE architecture uses interleaved attention layers with no positional encoding. Most layers use standard RoPE; the interleaved ones generalize beyond the 256K training window at inference via attention temperature scaling. Meta validated it with needle-in-haystack tests at 10M tokens and describes the long-term goal as "infinite" context.

API pricing across providers

Meta's direct Llama API is still waitlisted as of March 2026. Third-party providers are the practical option. There is meaningful price variation, and the available context window differs by provider even for the same model.

Llama 4 Scout (17Bx16E, 109B total)

Llama 4 Maverick (17Bx128E, 400B total)

Groq hosts Scout only (not Maverick). Together AI has Maverick on serverless but Scout is absent. Replicate uses time-based billing so direct token comparisons are not practical. Prices retrieved March 26, 2026. See our pricing page for updates.

What the benchmarks show

Meta's official model card compared Scout and Maverick against Llama 3.3 70B and Llama 3.1 405B, not GPT-4o directly. The release blog makes the GPT-4o claim in text. Here are the actual numbers from the model card, which is what you can verify.

Maverick's 69.8 GPQA Diamond and LMArena ELO of 1417 (from the "experimental chat version" per Meta) put it in the same range as GPT-4o. Scout's scores are solid for a budget model but clearly one tier below. The 20-point LiveCodeBench gap between them (32.8 vs 43.4) is worth noting if coding is your main use case. Independent testing by the LMArena leaderboard team placed Maverick's experimental version at 1417 ELO - above the standard GPT-4o but below GPT-4o-mini-high and the current frontier.

Image source: Meta AI

Self-hosting cost math

Scout (109B) fits on a single H100 80GB with int4 quantization (~54GB VRAM). Maverick with FP8 quantization needs a full DGX host: 8x H100s with 640GB combined VRAM. The hardware requirements are very different.

The breakeven numbers are large. 32 billion input tokens per month is roughly 1 billion tokens per day - sustained traffic of about 12,000 tokens per second. Maverick's 136B token breakeven is even further out of reach for most applications.

Self-hosting makes sense when you have data residency requirements, need Scout's full 10M context that providers cap, or are genuinely at hyperscale token volumes. Saving money at 10M tokens per month is not a realistic reason; the API wins by a large margin at that scale.

Scout's 10M context: the provider gap

Scout's native 10M context is real, but most providers do not expose it. Groq caps at 128K. DeepInfra at 320K. Fireworks at approximately 1M. The full 10M is only available when you self-host.

For long-document work, Scout at $0.08/M input is still the cheapest option by a significant margin. Gemini 3.1 Flash-Lite charges $0.25/M for 1M context. Claude Sonnet 4.6 charges $3.00/M (Anthropic removed the long-context surcharge earlier this year). Scout undercuts both, even capped at 320K.

If you genuinely need multi-million token context and cannot self-host, Gemini 3.1 Flash-Lite at $0.25/M is the most practical hosted option right now. Scout's provider caps make the 10M context more of a self-hosted research feature than an operational one for most teams.

Which one fits your workload

Neither model supports audio. Both support text and images natively (trained on both from pretraining, not added later). Knowledge cutoff is August 2024 for both. Worth keeping in mind for anything time-sensitive without a retrieval layer in front.

The license is not Apache 2.0

This comes up constantly. Llama 4 uses the Llama 4 Community License. For most developers and companies it works like a permissive commercial license: you can use it commercially, fine-tune it, and redistribute derivatives.

The restriction is the 700M monthly active user threshold. Platforms above that size need a separate license from Meta. That exists to keep WhatsApp-scale consumer deployments under Meta's terms. For developer tooling, APIs, or enterprise applications, it is not a practical constraint.

Sources

• Meta AI: Introducing Llama 4 (April 5, 2025)
• Meta Llama 4 Model Card - benchmark data and architecture details
• HuggingFace: Llama 4 Scout 17B-16E Instruct
• HuggingFace: Llama 4 Maverick 17B-128E Instruct
• Groq pricing
• DeepInfra pricing
• Fireworks AI pricing
• Together AI pricing