What this compares
Every row is a model you can call through an API today: Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral, Moonshot, Z.ai and MiniMax directly, plus open-weight models at the cheapest hosted price OpenRouter lists. For each one you get input, output and cached-input prices per million tokens, context window, maximum output, tool calling, structured output, reasoning and vision support, and the Artificial Analysis index where one is published.
The column that matters is the last one: your monthly cost. Per-token prices are hard to compare in your head because input and output are priced differently and most workloads are input-heavy. Enter your own token counts and request volume, and the table re-sorts by what you would actually pay.
How the monthly cost is calculated
For each model: requests × (uncached input × input price + cached input × cached price + (output + reasoning tokens) × output price). Reasoning tokens only count for models that support reasoning. If a provider charges more above a prompt-length threshold (Gemini above 200k tokens, for example), the higher tier applies to the whole request, as the providers bill it. Tick Batch API pricing to use the asynchronous batch rate where the dataset lists one.
Two things are not included: cache-write surcharges (Anthropic bills cache writes at 1.25× input) and tokenizer differences. The same text can be a different number of tokens on different models, so treat close calls as ties.
A worked example
A RAG endpoint doing 100,000 requests a month, 8,000 input tokens and 600 output tokens each, with 20% of the prompt cached. At October 21, 2026 list prices (this example recalculates when prices change):
- Claude Sonnet 5.5 ($2 input, $10 output, $0.10 cached per million): 640M fresh input × $2 = $1,280, 160M cached × $0.10 = $16, 60M output × $10 = $600. $1,896 a month.
- GPT-6 Luna ($0.10 input, $0.50 output, $0.01 cached): $64 + $1.60 + $30 = $95.60 a month, about 20× less.
Whether the cheaper model is good enough is a separate question. The model finder weighs that, and nothing replaces running your own eval on your own data.
Why there are two sources
Prices come from LiteLLM’s model price list, which is maintained by the people who build a widely used LLM gateway and tracks direct-provider prices. Every day the tool also pulls OpenRouter’s public model list and compares. When both list the same input and output price, you are looking at two independent confirmations. When they differ, the row is flagged sources differ and shows both numbers. It usually means a time-of-day discount (DeepSeek’s off-peak rates) or that OpenRouter is routing to a third-party host. Either way, check the official page linked on every row.
The table also remembers what prices looked like on your last visit, on this browser only, and flags new or repriced models when you come back.
Related
- AI Agent Cost Simulator: per-call prices understate agent costs because context grows every step.
- The Gate Costs More Than the Writing: a measured breakdown of where an agent’s API spend went.
- LiteLLM budget limits: max_budget needs Postgres.
FAQ
Which LLM API is cheapest?
It depends on your input-to-output ratio and how much of your prompt can be cached, which is why the table sorts by your workload rather than by list price. For short, output-light requests, small models from OpenAI, Google and the open-weight hosts usually come out cheapest. Filter by the capabilities you need first — tool calling and structured output remove a lot of the cheapest options.
What is a cached input token?
Most providers discount input tokens they have seen recently in an identical prompt prefix: a long system prompt, tool definitions, or a document you ask several questions about. Cached reads typically cost 10% or less of the normal input price. The “cached share” slider sets what fraction of each prompt you expect to hit the cache.
Are reasoning tokens billed?
Yes. Reasoning (“thinking”) tokens are billed at the output rate even though you usually do not see them. If you use a reasoning model, add an estimate in the reasoning-tokens field; it often exceeds the visible output.
How often are prices updated?
Once a day. The date at the top of the page is the last successful check, and the “What changed” panel lists price changes and new models as they are detected.
Why do open-weight models show a hosted price?
Open-weight models can be run by many hosts at different prices, or on your own hardware. The table shows the cheapest price OpenRouter lists, which is a useful floor. If you run the model yourself, your cost is your hardware and electricity; the model finder can filter by the GPU memory you have.
Can I download the data?
Yes. Download CSV exports the current filtered view with your workload costs. The full normalized dataset is available as JSON at /wp-json/dp-tools/v1/models.