LLM Token Cost Calculator
Compare what a real workload costs on Claude, GPT, Gemini, Nova, and open models — with prompt caching, batch discounts, and per-conversation unit economics. Includes an agentic budget-overrun simulator.
Workload
Start from a shape, then adjust anything.
Volume
Not sure how many tokens? Paste a sample prompt
Optimizations
The two levers that change the bill without changing the model.
Unit economics
Cost per conversation or agent run — the number a CFO can use.
Models
Pick up to 6 to compare.
| Model | Per request | Per unit | Per month | Per year | vs cheapest |
|---|---|---|---|---|---|
GPT-5 mini $0.250 in / $2.00 out per 1M Cheapest | $0.00080 | $0.016 | $1,217.60 | $14.6K | — |
Gemini 2.5 Flash $0.300 in / $2.50 out per 1M | $0.00099 | $0.020 | $1,506.78 | $18.1K | +$289.18 |
GPT-5 $1.25 in / $10.00 out per 1M | $0.00400 | $0.080 | $6,088.00 | $73.1K | +$4,870.40 |
Claude Sonnet 4.5 $3.00 in / $15.00 out per 1M | $0.00690 | $0.138 | $10.5K | $126.0K | +$9,284.20 |
Right-sizing opportunity. Moving this workload from Claude Sonnet 4.5 to Claude Haiku 4.5 would save $7,001.20/month (67%). Routing high-volume, low-complexity traffic to a cheaper model in the same family is usually the largest single lever available.
Where the money goes
Output tokens are priced far above input on most models — this is usually where a surprising bill comes from.
12-month projection
Top 3 cheapest options at a constant monthly growth rate.
How this calculator works
Everything runs in your browser against a hand-maintained price catalog — no account, no upload, no API keys. Three assumptions are worth knowing because they are where most calculators quietly cheat:
- Cache writes are charged. Cache misses pay the write rate, not the base input rate, on providers that price writes at a premium. That is why caching can show as a net loss at a low hit rate — because it is one.
- A month is 30.44 days, not 30. Over a year the difference is about two weeks of spend.
- Batch and caching are applied as independent shares of traffic. Providers differ on whether the two compose, so treat a configuration using both heavily as an optimistic bound.
Frequently asked questions
How do I calculate LLM API costs?
Multiply your average input tokens by the model input rate and your average output tokens by the output rate (both quoted per million tokens), add them for a per-request cost, then multiply by request volume. Output tokens usually cost three to five times input tokens, so an output-heavy workload costs far more than the headline input price suggests. This calculator does all of that plus caching and batch discounts.
Does prompt caching always save money?
No. Providers that charge a premium to WRITE to the cache (Anthropic charges 1.25x base input) make caching a net loss below a break-even hit rate of roughly 22%. Providers that cache automatically with no write premium, such as OpenAI, have a break-even of zero, so caching always helps there. This calculator flags when your configured hit rate is below the break-even for a given model.
What is the batch API discount?
Most major providers offer roughly 50% off for requests submitted through an asynchronous batch endpoint with a delayed completion window, typically 24 hours. It applies to any workload that does not need an immediate response: summarization, enrichment, evaluation, and offline classification.
Why do agentic workloads cost so much more?
An agent re-sends its accumulated context on every step of a tool-use loop and retries failed steps, so one user-visible task becomes many model calls with a growing prompt. Reported overshoot against pre-agentic budget plans clusters around 2.4x, which is why this calculator includes an explicit agentic toggle rather than leaving it to be discovered on an invoice.
What is cost per conversation and why does it matter?
Cost per conversation, per agent run, or per document is the unit economic that connects AI spend to business value. A total monthly figure cannot tell you whether a feature is profitable; a cost per unit compared against the revenue or saving per unit can.
This is an estimate. Finitizer shows your actual token spend.
TokenOps tracks real token usage by model, team, and feature across Anthropic, OpenAI, and Google — with budgets, anomaly alerts, and the same right-sizing analysis run continuously against what you actually spend, not what you estimated.