--- title: LLM Cost & Latency Calculator description: Estimate cost per request, monthly spend, response time and average concurrency for an LLM workload from tokens, caching, TTFT, generation speed and traffic. keywords: LLM cost calculator, LLM latency, TTFT, tokens per second, LLM API cost, prompt caching, LLM concurrency hide: - toc - navigation - footer ---
Tools · LLMs · 01

LLM cost and latency in one scenario.

Enter request size, workload and response characteristics. The calculator keeps price, response time and capacity separate so you can see which variable is actually limiting the system.

Cost modeltokens × rate
Latency modelTTFT + generation
Capacity modelLittle's law

Model and pricing

Presets use verified public pricing; every rate remains editable.

One request

Measured time to first token.
Generation rate after the first token.

Monthly load and limits

Cost / request — input + cache + output
Monthly cost — —
Response time — estimated time to the last token
Average concurrency required — average load; queues and bursts are not modelled
—
—
—
Where the cost comes fromper request
Uncached inputCached inputOutput
Where the time comes fromcomplete response
TTFTGeneration after first token
TTFT—
Generation—
Approximate capacity—
Capacity headroom—
Method

What it calculates — and what it does not.

Cost. Uncached input, cached input and output tokens are priced separately. Preset rates are a starting point: replace them with your contract, batch pricing or effective rate whenever that better represents your workload.

cost = (Tin,uncached × Pin + Tin,cached × Pcache + Tout × Pout) / 1,000,000

Latency. TTFT covers everything up to the first generated token. The remaining generation is approximated as (output_tokens − 1) / tokens_per_second. This estimates completion time, not perceived streaming latency.

response_time ≈ TTFT + (Tout − 1) / generation_rate

Capacity. For a stable workload the calculator applies Little's law: average concurrency ≈ arrival rate × service time. This is useful for average capacity planning; it is not a queueing model for bursts, p95/p99 latency, provider rate limits, retries or batching.

average_concurrency ≈ requests/second × seconds/request

Tool calls, web search, cache storage, audio/image charges, priority processing, volume discounts and self-hosted infrastructure are excluded unless you fold them into the editable rates. Presets with known long-context pricing rules apply them automatically and disclose the adjustment next to the source.

01What is an LLM?Tokens, generation and context before converting them into cost.Concept 02Latency, streaming and human interactionWhy TTFT, generation rate and completion time describe different user experiences.Chapter 03Voice-agent architecturesHow latency budgets compound when the LLM is only one stage in a realtime system.Engineering

Preset sources

The data layer stores the source organization, primary URL and verification date for every preset. This version uses official pages from OpenAI, Anthropic and Google AI for Developers. Prices can change; check the source shown for the selected model before making a contractual decision.