LLM API price table
All prices in USD per 1M tokens. Updated daily from OpenRouter's public API. Type to filter; click a header to sort.
| Model | Provider | Input $/1M | Output $/1M | Context | Blend* |
|---|
*Blend = (3×input + 1×output)/4, a rough proxy for a typical assistant workload. Cheapest useful models for bulk work are not always the ones with the lowest input price — check the blend. Models that return images, audio or video are listed here but are deliberately kept out of the calculator dropdowns: they bill on a different axis and cannot be compared per token.
Embedding models
Embedding endpoints bill on input tokens only. They are not listed by the main model API, so this table comes from a second public source (LiteLLM's community price list) — the same list the RAG and embedding calculators read.
| Model | Provider | Input $/1M | Max input |
|---|
How to read this table
- Input and output are billed separately, and output is usually several times the price of input. A model that looks cheap on the input column can be expensive for workloads that generate long responses — chat, summarisation, code generation.
- Context window is a capability, not a price. It tells you what the model can accept, not what it costs. Long-context requests may carry a higher rate above a size threshold, and an oversized window makes it easy to send context the task never needed.
- Blend is a sorting aid, not an estimate. It weights three parts input to one part output. That is a reasonable proxy for an assistant re-sending a lot of context, and a poor one for a pipeline that mostly generates.
- Some listings are excluded from the calculators on purpose. Models returning images, audio or video bill per output unit rather than per token, and batch or priority variants carry a different rate. Both are visible here, both stay out of the dropdowns.
The spread is wider than people assume
Placed end to end, per-token prices on this table span roughly three orders of magnitude. That enormous range is genuinely useful — it means model choice can move a bill by multiples — but it is only true for workloads where token price dominates. On a voice product, in a retrieval system, or anywhere operational overhead is significant, a hundredfold cheaper model can still move the total by a few percent.
The table answers "what does this model charge". It does not answer "what will my system cost". Those two questions are much further apart than they look.
About the embedding table
Embedding endpoints bill on input tokens only — there is no generated text to charge for — which is why the price column here has no output counterpart. Two things matter more than the headline rate when you are budgeting embeddings: how text is chunked, since overlap can inflate indexed tokens by ten to fifteen percent, and how many dimensions the model produces, because higher-dimensional vectors cost more to store and search even when the per-token price is identical.
Choosing a model from this table
The cheapest row is rarely the right answer, and never the whole answer. Three questions narrow the field faster than any ranking.
- Is your bill input-heavy or output-heavy? Chat, summarisation and code generation are output-heavy, so the output column decides. Retrieval, classification and extraction are input-heavy, so the input column does.
- Does the task need reasoning, or retrieval? Most production traffic is lookup and reformatting, which small models handle at a fraction of the price. Reserve the expensive tier for requests that genuinely need it, and route between the two.
- What does a wrong answer cost? A bad marketing draft is an inconvenience; the same mistake in a regulated workflow is not. Paying for the larger model on the small share of high-stakes calls is usually cheaper than paying for it on all of them.
Worked in that order, model selection stops being a race to the bottom of a price column and becomes a routing decision — which is what it actually is.