How do I calculate LLM API costs?

Cost per call = (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate). Multiply by calls per day × 30 for a monthly figure. A 6,000-in / 500-out RAG answer on Claude Sonnet 4.6 costs about a cent; the calculator on this page does the arithmetic against live catalog rates.

How many tokens is a typical request?

Rules of thumb: 1 token ≈ 4 characters of English (≈ 0.75 words). A short support reply runs ~800 input / 200 output tokens; a RAG answer with retrieved context ~6,000 / 500; one coding-agent step ~10,000 / 1,500; a long-document analysis can be 50,000+ input tokens. CJK languages consume roughly 1.5–2× more tokens per character.

Why is output more expensive than input?

Generation is sequential and compute-heavy, so providers price output tokens ~4-5× input. Practical consequence: capping max_tokens is one of the highest-leverage cost controls — many workloads over-generate by default.

How accurate is this calculator?

The per-token rates are Kunavo's live catalog prices — the exact numbers the API bills, shown next to each provider's official list price. Real spend can be lower with prompt caching (cached input bills at ~10% of the input rate) and failed requests are never billed.

LLM API Cost Calculator — Claude, Gemini & GPT token pricing (live rates)

Every LLM bill reduces to one formula: (input tokens ÷ 1M × input rate) + (output tokens ÷ 1M × output rate), summed over your calls. The calculator below runs that math against Kunavo's live catalog rates — the same numbers the API actually bills, listed next to each provider's official price — so you can budget a workload for Claude, Gemini and GPT in seconds.

Model

Input tokens / call

Output tokens / call

Calls / day

Per call

$0.00092

Per day (500 calls)

$0.458

Per month (30d)

$13.73

saves $32.02/mo vs the provider's list price ($45.75)

Rates are Kunavo's live catalog prices (USD per 1M tokens), the same numbers the API bills. Prompt caching, batch discounts and per-model nuances can lower real spend further.

Reading the results

The presets encode realistic token shapes: a support reply is light (~800 in / 200 out), a RAG answer carries retrieved context (~6,000 in / 500 out), one coding-agent step is heavier still (~10,000 in / 1,500 out), and long-document analysis is input-dominated (50,000+ in). If you don't know your own shape yet, log usage.prompt_tokens and usage.completion_tokens from a day of real traffic — every OpenAI-compatible response carries them.

Worked examples at current rates

Workload	Shape (in / out)	Model	Per call	1,000/day → per month
Support bot	800 / 200	`claude-haiku-4-5`	$0.00072	$21.60
RAG assistant	6,000 / 500	`claude-sonnet-4-6`	$0.0102	$306
Classifier / router	400 / 20	`gemini-2-5-flash`	$0.000051	$1.53

Full per-model rates live on the pricing page — every model shows its Kunavo price next to the provider's official rate (typically 60%+ under list for Claude).

Four levers that shrink the number

Prompt caching. A large stable system prompt or knowledge block bills at ~10% of the input rate once cached — for prompt-heavy apps this alone halves the bill. See the caching deep dive.
Model tiering. Route easy calls to gemini-2-5-flash or claude-haiku-4-5 and reserve frontier models for the hard fraction — the calculator makes the delta obvious.
Cap max_tokens. Output is ~5× input; unbounded generation is unbounded spend.
Trim retrieval. Fewer, better-ranked RAG chunks cut input tokens without hurting quality. Patterns in the cost optimization guide.

FAQ

How do I count tokens before sending?

Estimate with 1 token ≈ 4 English characters, or count exactly: Anthropic exposes a count-tokens endpoint and OpenAI-wire models tokenize with tiktoken. For budgeting, yesterday's usage fields beat any estimator.

Do failed requests cost money?

On Kunavo, no — failed requests are never billed, and the prepaid wallet (from $5, never expires) fails closed at zero instead of surprise-invoicing you.

Can I get these rates programmatically?

Yes — GET /v1/models lists every enabled model, and llms.txt carries a machine-readable catalog snapshot for agents. Start with a free account and the quickstart.