Free calculator · dated rate card · no signup
Token cost calculator for every AI user and call.
Use real calls and token counts to calculate cost per user, cost per call, and total model cost. Cache reads, cache writes, and retry attempts are priced separately.
How much do 1,000 tokens cost?
Divide the model's price per million tokens by 1,000. At this calculator's recorded GPT-4.1 mini rates, 1,000 uncached input tokens cost $0.0004; 1,000 output tokens cost $0.0016. A call with both costs $0.002. Input, output and cache tokens have different rates, so there is no single universal token price.
| Model / token type | Per 1,000 tokens | Per million tokens |
|---|---|---|
| GPT-4.1 mini · uncached input | $0.0004 | $0.40 |
| GPT-4.1 mini · cache read | $0.0001 | $0.10 |
| GPT-4.1 mini · output | $0.0016 | $1.60 |
| Claude Sonnet 4.5 · uncached input | $0.003 | $3.00 |
| Claude Sonnet 4.5 · output | $0.015 | $15.00 |
Try the 1,000-input-token example, then replace its sample counts with your own. Verify current prices in the OpenAI or Anthropic rate card; batch, long-context, tools and contract-specific charges may differ. This page estimates model usage, not your complete application bill.
Enter one period of model usage
All model rates and results are in USD.
How to read it
Why cache and retries stay separate
Cached input is usually billed below standard input. Counting it at the full input rate overstates cost. This calculator keeps uncached input and cache reads as separate, non-overlapping buckets. Provider-reported cache writes are added at their own rate, and retries increase the number of paid attempts.
Calculation method
Cost per attempt = (uncached input tokens × input rate + cache-read tokens × read rate + supported cache-write tokens × write rate + output tokens × output rate) ÷ 1,000,000. Token buckets are separate: do not count cache reads again as uncached input.
Attempts = active users × calls per user × (1 + extra retry percentage / 100). Total AI cost = attempts × cost per attempt. AI cost per active user = total AI cost ÷ active users. A 10% retry rate means 10 additional attempts per 100 base calls, not a 10% probability of unlimited retries.
Worked example · synthetic, USD
For Claude Sonnet 4.5: 1,000 uncached input tokens at $3, 500 cache-read tokens at $0.30, 200 five-minute cache-write tokens at $3.75, 100 one-hour cache-write tokens at $6, and 300 output tokens at $15 per million cost $0.009 per attempt. Ten users × 20 calls × 1.10 = 220 attempts: $1.98 total, or $0.198 per user.
This measures estimated model cost, not net profit. Users and calls must be positive integers. It excludes taxes, hosting, provider credits, batch discounts and unentered tool or network charges. Averages can hide heavy users; inspect retry pressure before applying the result to a paid plan.
Price basis and limits
Averages are not account-level attribution. Read how to measure AI cost per user, or use cost per successful agent task when failures still consume tokens.
These are dated catalog rates, not a live price feed. OpenAI and Anthropic entries were recorded on ; BytePlus Seed 2.0 Lite on . Check the selected model's source and verification date shown with the result before budgeting. Re-running this page does not update those dates.
Only supported token buckets are priced. Models without a separate cache-write rate reject cache-write input. Seed 2.0 Lite rejects more than 128,000 input-side tokens per request because the higher pricing tier is not modeled. No FX conversion is applied.
Provider references: OpenAI pricing, Anthropic pricing, and how Grow or Die versions prices. This tool is an estimate, not a reconciled provider bill.