Products

What does your LLM workload really cost?

Model your monthly LLM API spend in the browser — then see how much prompt caching, batch processing and routing part of your traffic to a cheaper model would save. Compare 18 models from 6 providers on your exact workload.

Monthly & yearly Live recompute Official prices No signup
Workload template

Templates are example workloads — starting points to adjust to your own numbers.

Model
Traffic

Tokenizers differ between model families — measure on your own data. Why token counts vary.

Prompt caching
5 min (1.25× write)
1 hour (2× write)
Batch
Routing to a cheaper model
Forecast & currency
Enter your traffic on the left to see the estimate.

Optimized spend

at baseline

Per year

optimized, at set growth

Per task

Where the savings come from

Levers apply in order: caching, then batch, then routing. A greyed lever can't apply to this model or workload.

Compare models on this workload

Every model priced on your exact traffic — baseline and optimized (caching + batch). Sorted by optimized monthly cost.

ModelBaseline / moOptimized / moSavingPer task

12-month projection

Optimized monthly spend compounded at your growth rate.

How to get there

Get the three highest-impact recommendations for your workload — and an emailed PDF of this estimate (breakdown, model comparison and projection).

    How the estimate works

    A transparent, token-level model — no black box. Every figure recomputes live as you change an input, and you can see the exact numbers behind it.

    Frequently asked questions

    How is LLM API cost calculated?
    Cost is driven by tokens, not requests. We multiply your monthly calls by the average input and output tokens per call, price input and output separately at each model's per-million rate, then apply caching, batch and routing savings. The result is API spend only — monthly and yearly.
    How much does prompt caching save?
    Caching helps when a large, stable prefix repeats across calls. Cached reads are typically ~10% of the input price, so the saving scales with your cacheable share and hit rate. If the cacheable prefix is shorter than the model's minimum, caching doesn't apply and the saving is zero.
    When should I use batch processing?
    Batch APIs usually give a 50% discount for asynchronous, higher-latency processing. Use it for work that doesn't need an immediate answer — extraction, classification, offline generation. Not for interactive chat.
    Is a cheaper model always cheaper per task?
    Not always. A cheaper model may need more calls, longer prompts or retries, and its caching/batch support differs. The comparison table prices every model on your exact workload, so you see real per-task cost, not the sticker rate.
    How often are prices updated?
    Prices come only from official provider pages, refreshed monthly and on new-model releases. Each model carries a source link; the reference file carries an updated date, and if it's older than 60 days the calculator shows a notice.
    Do token counts and estimates include infrastructure?
    The estimate covers LLM API token spend only — it excludes servers, vector databases and embeddings, fine-tuning, monitoring and engineering time. Also note tokenizers differ between model families, so the same text can be more or fewer tokens on another model; measure on your own data.

    Take it further

    GenAI Integration →Ship LLM features with cost, latency and quality under control. LLM Fine-tuning →When a tuned smaller model beats prompting a big one — we benchmark it. Fine-tuning cost benchmark →Real numbers on when fine-tuning pays back. Claude vs GPT for product teams →Cost, caching and batch, compared for real workloads. AI chatbot development cost →What a production chatbot really costs to build and run.
    Link copied