LLM API cost calculator
Pick a model, put in the input and output tokens you get through in a month, and this works out what that month costs in 5 places at once. Output tokens are billed at the output rate, not the input rate — that distinction is worth several times the bill on a chat workload. Each row shows the unit prices it used, so you can check the multiplication rather than take it on trust.
What that month costs
USD / 1M tokens
| Where you call it | Per month | Unit price |
|---|---|---|
| Y-APIYou are here | Per month$0.105 | Unit price$0.0075 in · $0.015 out per 1M |
| OpenRouterdeepseek/deepseek-v4-flash-0731 | Per month$0.843dated snapshot build, not the rolling model | Unit price$0.06 in · $0.12 out per 1M |
| DeepInfradeepseek-ai/DeepSeek-V4-Flash | Per month$1.264standard tier; priority and flex tiers differ | Unit price$0.09 in · $0.18 out per 1M |
| Together AIDeepSeek V4 Flash 0731 | Per month$1.963dated snapshot build, not the rolling model | Unit price$0.14 in · $0.28 out per 1M |
| DeepSeekMost expensivedeepseek-v4-flash | Per month$3.521,2off-peak rate; a cheaper cache-hit tier also exists | Unit price$0.22 in · $0.66 out per 1M |
The dearest of the priced rows costs $3.52 — 33.5× this bill.
Priced against 4 of the 4 places we check for this model. Rows without a figure say which of the two reasons applies.
The 20:1 top-up rate is limited-time. At the regular 1:10, the same month costs $0.21 here. Credit already in your account is not repriced.
- 1off-peak rate
- 2a cheaper cache-hit tier also exists
- 3dated snapshot build, not the rolling model
- 4standard tier; priority and flex tiers differ
Method
How this bill is worked out
- Input and output priced separately
- Output tokens cost two to four times input tokens on every provider in this table. A calculator that bills all tokens at the input rate understates a chat workload by a wide margin, so both counts are asked for and both rates are applied.
- Our figure is cash, not credit
- Our rates are quoted in USD credit, and a $1 top-up currently becomes $20 of credit. Our row divides by 20 to get the dollars that actually leave your card, which is the only unit the other rows are in.
- Their cheapest published tier
- Where a provider publishes several rates, this takes the cheapest one that applies to an ordinary real-time call — off-peak rather than peak, standard rather than priority, cheapest region, shortest input-length tier, after any limited-time discount. Cached input is excluded everywhere, including ours. Every such choice is marked under the table and every one of them favours the other provider.
- Nothing is estimated
- A row either has a price read off that provider's own pricing page, or it says so. No averages, no "approximately", no figures from third-party posts — for one DeepSeek model alone, three widely-cited posts quote three incompatible prices.
Limits
What this number does not include
Prompt caching
Vendor APIs bill cached input far below their cache-miss rate and we have no equivalent tier. If you replay a long, stable system prompt thousands of times, the vendor can come out cheaper than us — this calculator will not show that, because it prices every token as a cache miss on both sides.
Reasoning tokens
On reasoning models, tokens that are billed but never appear in the response are not in your output count. One probed model billed 21.8× its output cap. If you are calling a reasoning model, treat the output figure you enter as a floor.
Everything that is not tokens
No provider in this table charges a monthly fee, so the bill is tokens only. It also means the arithmetic here says nothing about rate limits, latency or which models a provider actually has — and those are what usually decide the choice.
FAQ
Questions about this calculation
- Why do I have to enter input and output tokens separately?
- Because they are priced differently everywhere, usually by a factor of two to four. A single "total tokens" box would force an assumption about your input-to-output ratio, and the bill would be wrong by roughly that factor for anyone whose workload does not match it. The default scenario shown on load is 5:1, which is close to a typical chat or RAG load — your own log will tell you your actual ratio.
- Why is Y-API's bill so much lower?
- Almost entirely the top-up rate. Read our credit rates as if they were dollars and they sit roughly on top of the vendors' own list prices; the gap you see is the 20:1 rate at which a dollar becomes credit. That rate is limited-time and reverts to 1:10, which doubles our figure — the line under the table shows what that month would cost.
- Why are some rows empty?
- Two different reasons, and the table never merges them. One says the provider does not carry that model at all. The other means we have not verified a price on their own pricing page — not that the model is unavailable, and not that it is cheap. Filling either in from a secondary source would make the total look complete while making it wrong.
- How current are these prices?
- The external figures were each read off that provider's own pricing page; the oldest of those readings is 2026-08-26, which is the date shown at the top. Our own rates come from the live catalog at build time, so they are never staler than the last deploy. Providers do change prices without notice, so for a decision that matters, open the source link and check.
Provenance
Where the prices come from
Every external figure in this calculator comes from the same dataset as the full market comparison, which lists one source link per provider alongside the date it was read.
Check the arithmetic against a real request
The fastest way to verify any of this is one call. Sign in, send a request, and the usage log shows the exact token counts and the exact deduction — which is the same multiplication this page just did.