benchrPricing options
| Option | Cost |
|---|---|
| Self-hosted (Meta weights) | Infrastructure only |
| Together.ai input | ~$0.11/1M |
| Together.ai output | ~$0.34/1M |
| Context window | 10,000,000 tokens |
10M context: what it actually unlocks
Ten million tokens is approximately 7,500 pages of text. To put this in engineering terms: a 1M-line codebase with documentation fits comfortably. A decade of customer support transcripts. The complete works of a prolific author many times over. The second-largest commercial context window — Gemini 3.1 Pro at 2M — handles 1,500 pages. Scout handles 5× that. For applications where the constraint has always been context length — not reasoning quality — this is a fundamental capability unlock.
Practically: a software company could load its entire 800K-line production codebase plus 200K lines of tests and documentation into a single Scout context call. No chunking, no retrieval, no loss of cross-file reasoning from splitting. This is not a marginal improvement over 2M context — it enables workflows that were architecturally impossible at smaller context sizes.
Scout vs Maverick: the context-quality tradeoff
Llama 4 Maverick is the stronger reasoning model — better on SWE-bench, deeper on complex analysis, more capable on general benchmark tasks. Scout trades reasoning depth for context length and throughput. The decision is binary based on your context requirement: if your task fits in 1M tokens, use Maverick for better quality at slightly higher cost. If your task requires 2M–10M tokens, Scout is the only option — there is no alternative with this context range, open or closed.
Hardware for self-hosting
Scout's MoE architecture is lighter than Maverick, with fewer active parameters per inference step. Full precision fits on approximately 2× A100 40GB; with quantization, a single A100 80GB or H100 handles serving at reasonable throughput. For the 10M context window to be practically usable, you need sufficient KV cache memory — processing a 10M token context call requires substantial VRAM. In practice, most self-hosted deployments target the 1M–2M context range where memory requirements are more manageable.
Cost scenarios
At 5M input + 1M output per month (long-document processing): approximately $0.55 + $0.34 = ~$0.89/month via Together.ai. Gemini 3.1 Pro at the same volume (inputs under 128K): $10 + $12 = $22/month — 25× more expensive. The Scout economics are so favorable that the question becomes: is the quality gap on your specific task worth the 25× price difference? For tasks specifically requiring 2M+ context, there is no Gemini alternative — Scout is the only viable option.
Use-case fit
Best for: Full-repository code analysis; multi-year document archive processing; long-running agent sessions that must maintain complete context without summarization; any workflow where the 2M context ceiling of the next available option is a hard constraint; cost-sensitive long-context retrieval where chunking introduces coherence loss.
Skip if: Your tasks fit in 1M context and reasoning quality is more important than context ceiling — Llama 4 Maverick or Gemini 3.1 Pro are stronger reasoning models. Also skip for multimodal (image) tasks — Scout is text-only.
Decision checklist
Identify your actual p95 context length requirement. If it's under 500K tokens, Llama 4 Maverick (1M context, stronger reasoning) is the better choice. Only choose Scout when you have tasks that regularly need 1M+ tokens and the reasoning quality tradeoff is acceptable.
Verify provider support: contact your inference provider to confirm maximum supported context length before building architecture that depends on the full 10M window. Self-hosting with dedicated hardware is the reliable path to the full context advantage.