Methodology
Every number on LLM Cost Index comes from the public OpenRouter models API, collected automatically once a day.
Tracking since Oct 21, 2026 · 1 daily snapshots · last Oct 21, 2026
What we collect
- Every listed model: prices per token (input, output, cached input, long-context tiers), context window, maximum output, input/output types, tool calling, reasoning, open weights, announced retirement dates and the Artificial Analysis intelligence index when available.
- Every provider that serves each model, with its own prices, service tier and quantization.
How we calculate
- Prices
- Converted to US dollars per 1 million tokens. Prices can differ from a provider’s own API; always confirm before committing.
- Ranking by price
- A 3:1 mix of input to output tokens, the typical shape of chat traffic.
- Monthly cost
- Requests × 30 days × (uncached input × input price + cached input × cached price + output × output price). Workloads used on model pages: customer chatbot (1,000 conversations a day, 8 turns each); document q&a (rag) (500 questions a day over retrieved documents); coding / tool-using agent (200 tasks a day, long tool-heavy contexts); content generation (30 long articles a day); classification / extraction (100,000 short items a day).
- Cheaper alternatives
- Models with the same tool calling, image input and structured output, enough context, and at least 20% lower blended price, most capable first.
- Provider uptime
- OpenRouter publishes each provider endpoint’s uptime over the last 30 minutes and the last day. We sample it four times a day and average the daily figures over 30 days. A provider’s “bad day” is a day its endpoints averaged under 99%.
- Market price index
- The median blended price (3:1) of all paid models, and of three tiers by intelligence index (40+, 20–40, under 20), recorded with every daily snapshot. Free models, batch/flex variants and routers are left out.
- Price changes
- Compared with the previous daily snapshot; moves under 1% are ignored as rounding.
Articles
Our articles are written from the same daily data. Every number in an article comes from the snapshot, never from memory; an automatic check rejects any statistic that is not in the data, and each article ends with a note on how it was measured.