Free tool
LLM cost calculator.
Pick a model, enter your volume and token sizes, and see the monthly and annual cost — plus the chunk a cheaper tier, semantic caching, and a hard budget cap would reclaim.
Prices are indicative list prices — always confirm with the provider before budgeting.
Estimated spend
$2,661 /mo
$31,938/yr · $0.0088 per call · 10,000 calls/day on GPT-4o
Typically reclaimable, governed
$6,388–$12,775 /yr
A typical 20–40% of repetitive traffic comes back with routing + semantic caching, and a hard cap stops the spike that blows the quarter. An estimate, not a quote.
How the math works: cost = requests × (input tokens × input rate + output tokens × output rate) ÷ 1,000,000. Output tokens usually dominate — they cost several times more than input — so a chatty model on a high-volume feature is where bills run away.
Why the “reclaimable” number:repetitive production traffic (support, RAG, agents re-asking) is often 20–40% cacheable, and routing the easy calls to a cheaper tier saves more. A hard cap stops the runaway agent before it spends. It's an estimate — your real number depends on your workload, which the free audit measures directly.
See the underlying rates on the LLM pricing tracker →
Frequently asked questions
How do I calculate the cost of an LLM API call?+
Cost = (input tokens × input rate + output tokens × output rate) ÷ 1,000,000, where rates are the model's dollars-per-million-tokens. Multiply by the number of calls for a period. Output tokens usually cost several times more than input, so a chatty model on high volume is where bills grow fastest.
Why are output tokens more expensive than input tokens?+
Generating text is the computationally expensive part — the model runs a forward pass per output token. Providers typically charge 3–5× more per output token than per input token, which is why reasoning models (which generate large amounts of hidden 'thinking' output) can cost more for a short visible answer.
How can I reduce my LLM costs?+
Four levers compound: route easy requests to a cheaper model tier, cache repetitive traffic (semantic caching can remove a large share of production calls), compress prompts to cut input tokens, and put a hard budget cap in front of runaway agents. On typical repetitive production traffic, 20–40% is reclaimable.
What is a cache hit rate and how does it change my bill?+
The cache hit rate is the share of requests answered from a stored result instead of a fresh model call. A semantic cache hit costs nothing at the provider, so a 30% hit rate on repetitive traffic removes roughly 30% of that workload's provider cost. Slide the cache-hit control above to see the effect.
Are these prices exact?+
They're indicative list prices meant for estimation. Batch, volume, and enterprise-negotiated rates differ, and providers change prices. Always confirm with the provider before budgeting — or point the free AI Spend Audit at your real usage export to measure your actual numbers.