LLM cost calculator: what your API calls cost a month

How to use the LLM cost calculator

Start with the model you run today and the number of calls it takes in a month; your provider's dashboard or your own logs have both. Input tokens are everything you send on each call — the system prompt, the conversation so far, any retrieved documents and the user's message. Output tokens are the reply. Use averages: one unusually long call does not change a monthly bill, a long system prompt sent on every call does.

The last field is the one only you can fill in: the share of calls that end in a decision rather than a piece of writing. Classifying a ticket, choosing which tool or model to call, scoring an answer against a rubric, checking a reply before it ships — those are decisions. Leave it at zero to price the LLM alone.

Input tokens vs output tokens

Almost every model charges more for what it writes than for what it reads. In the comparison table, output costs a median 5 times the input price, so a long reply moves the bill more than a long prompt of the same length. Capping the reply length, asking for a label instead of a paragraph, and returning structured fields are usually the cheapest changes you can make to an LLM cost estimate before touching the model.

How many tokens is a request?

In English a token is about four characters, so 1,000 tokens is roughly 750 words, per OpenAI's guide to tokens. Other languages usually take more tokens for the same text. The surer way is to read it off your own traffic: most APIs report the input and output token counts in a usage field on every response, and a day of those numbers averaged is a better input for this LLM cost calculator than any estimate.

Watch for the costs hiding in the input. A chat app re-sends the whole conversation on every turn, so the twentieth message carries the nineteen before it; a retrieval pipeline adds every passage it fetched. Both grow the input side of the bill without anyone writing a longer prompt.

LLM cost per token: the price list

The models in the comparison, at their listed price per million tokens. Click any model in the LLM cost calculator above to price your own workload on it.

Where Jev lowers an LLM bill

Jev is a decision model: it reads text and returns a probability for each option you give it, and never writes a reply. That makes it no use for drafting an answer and very cheap for choosing one. At the defaults of the LLM cost calculator above — 1,000 input and 300 output tokens — one call costs $0.00025 on OpenAI: GPT-6 Luna and $0.010 on Anthropic: Claude Opus 5.5, against $0.000042 for the same input on Jev, which bills no output at all.

Cheaper is not the same as good. TypeSafe's own evaluation puts Jev at 67.8% against 67.9% for GPT-5.6 Terra, but it measures agreement with other frontier models rather than correctness, and our own side-by-side runs found cases where the LLM wins outright. Test the decisions you would move on your own data first. The usual pattern is a gate: Jev answers the clear cases and sends low-confidence ones on to the LLM, so the LLM cost falls without the hard cases getting worse.

Other ways to cut LLM cost

• Batch what can wait. 8 of the 13 models above have a batch listing on OpenRouter, at a median 50% of the standard input price.
• Cache the prefix. Providers that support prompt caching charge less for a system prompt or document they have seen recently, which helps most when every call starts the same way.
• Shorten the reply. Output is the expensive side, so a label or a short JSON object costs a fraction of a paragraph.
• Route by difficulty. Send easy calls to a small model and hard ones to a large one. TypeSafe offers this as Jev Router on OpenRouter, which picks the model and reasoning effort per request; the model routing use case shows the same idea with Jev called directly.

LLM cost calculator questions

Want the numbers for your own traffic rather than an estimate? Every call to this site's API returns its token count, and usage blocks from your current provider work the same way.