Put in your real traffic — tokens per call and calls per month — and see what 30 models from Anthropic, OpenAI, Google, DeepSeek, xAI, Mistral, Meta, Amazon and Cohere would actually cost you per month. Every price is editable, so you can drop in your own negotiated or provisioned rate. No signup, nothing stored.
That's input and output tokens a month.
| Model | In $/M | Out $/M | Blended $/M | Cost / month | vs cheapest |
|---|
Tap any price to edit it. Click a column heading to sort. Blended $/M is your effective cost per million tokens at your input:output ratio — the only per-token number worth comparing.
In that order of impact, for a typical retrieval or agent workload.
Most production traffic is easy. Sending every request to a frontier model is where 60–90% of the waste lives. Classify first, escalate only on low confidence, and the blended rate collapses toward the small-model row above.
System prompts, tool schemas and retrieved documents repeat on every call. Cached input reads cost roughly a tenth of fresh input on Anthropic and OpenAI. Set the cache slider above to your real hit rate and watch the input column fall.
Evals, backfills, enrichment, summarisation jobs and nightly classification do not need a synchronous response. The batch APIs are 50% off for exactly the same model.
For agents and scripts. No key, no signup.
curl https://smeltworks.com/llmcost/api/prices
curl "https://smeltworks.com/llmcost/api/estimate?input_tokens=4000&output_tokens=800&calls=100000&cached_pct=70&batch=false"The calculator tells you what you'd pay. The kit is how you actually get there: a drop-in cost-aware router, a token-accounting middleware, a prompt-caching checklist, and a spend model you can hand to finance.
There is no flat answer — it is entirely a function of tokens. A support bot doing 100,000 calls a month at 4,000 input / 800 output tokens is 480M tokens of input and costs roughly $1,400/month on a frontier model and under $60 on a small one. Put your own numbers in the calculator above.
Roughly 4 characters of English, or about 0.75 words. 1,000 tokens ≈ 750 words ≈ 1.5 pages of plain prose. Code and JSON are denser — budget nearer 3 characters per token.
Input tokens are processed in parallel in a single forward pass; output tokens are generated one at a time and each one occupies the accelerator. That is why output is typically 4–5× the input price, and why "make the model answer more briefly" is a real cost lever.
Yes. Reasoning tokens are billed as output even though you never see most of them. A model priced identically to a non-reasoning peer can still cost several times more per answer. Estimate output tokens generously for reasoning models.
They are hand-maintained and reviewed on 11 August 2026, with a link to each vendor's own pricing page in the JSON feed. Vendors change prices, add tiers and charge differently above certain context lengths. Treat this as a planning tool, verify before you commit, and edit any cell to match your contract.