Long-context models are everywhere now — nearly every flagship ships a 1M-token window, and a couple go far past it. What nobody puts in the marketing copy is how wildly the cost of using that window varies.
Filling a 1M-token context window once — just the input tokens you send — ranges from $0.05 to $30 depending only on which model you point at. That's a 600× spread for the same nominal capability.
Here's the whole ladder, GA models with a ≥1M-token window, sorted by what one full input pass costs:
| Model | Provider | Context | 1M-token input call | Output /1M |
|---|---|---|---|---|
| Qwen-Flash | Alibaba | 1M | $0.05 | $0.40 |
| Llama 4 Scout | Meta | 10M | $0.10 | $0.30 |
| Gemini 3.1 Flash-Lite | 1M | $0.25 | $1.50 | |
| Amazon Nova 2 Lite | Amazon | 1M | $0.30 | $2.50 |
| Grok 4.3 | xAI | 1M | $1.25 | $2.50 |
| Gemini 3.5 Flash | 1M | $1.50 | $9.00 | |
| GPT-5.4 | OpenAI | 1M | $2.50 | $15.00 |
| Claude Sonnet 5 | Anthropic | 1M | $3.00 | $15.00 |
| Claude Opus 4.8 | Anthropic | 1M | $5.00 | $25.00 |
| GPT-5.5 | OpenAI | 1M | $5.00 | $30.00 |
| Claude Fable 5 | Anthropic | 1M | $10.00 | $50.00 |
| GPT-5.5 Pro | OpenAI | 1M | $30.00 | $180.00 |
Why this matters more than the sticker price
For a lot of real workloads — RAG over a big corpus, whole-repo code questions, long-document summarization — you're mostly paying to read a large context, not to generate a long answer. In that regime the input rate dominates your bill, and the model choice moves it by two-and-a-half orders of magnitude.
Concretely: if your job stuffs ~800K tokens of retrieved context into every call, running it on GPT-5.5 Pro costs ~$24 per call in input alone; the same call on Qwen-Flash is ~4 cents. Whether that 600× gap is worth it depends entirely on whether the frontier model's answer quality actually changes your outcome — but you should at least know you're making that trade.
Two honest caveats
- This is input only. Output tokens are billed separately and, for the pricey models, are even steeper (GPT-5.5 Pro is $180/1M out). If your workload is generation-heavy rather than context-heavy, the math shifts.
- Cached input changes everything for repeated prefixes. If the big context is a static prefix you reuse — a fixed system prompt, tool definitions, a document you ask many questions about — cache reads bill at roughly a tenth of the standard input rate on the major providers (and up to ~50× cheaper on some DeepSeek tiers). A 1M-token context you query 20 times isn't 20 full-price reads; it's one, plus 19 cache hits.
The point
"1M context" is a capability checkbox that hides a 600× cost range. Before you pick a model for a long-context job, price the input pass at your real token volume — that single number usually decides the bill more than anything on the model card.
Prices are USD per 1M tokens, pulled from official provider pricing pages and updated daily. The live, sortable context-vs-price table (and a calculator for your exact token counts) is at aimodelwatch.dev.
Top comments (0)