LLM API Cost Calculator — compare 42 models in seconds

It's a BPE-approximate estimator that runs entirely in your browser — it mimics how modern tokenizers split words, numbers, punctuation and CJK characters, and is typically within ±10–15% of the real count for English prose. Each provider uses a slightly different tokenizer (o200k, Claude's tokenizer, SentencePiece variants), so even "exact" counts differ between models. For billing-critical work, use the provider's official token-counting endpoint.

Prices are standard pay-as-you-go API list rates in USD per million tokens, compiled from public provider pricing pages and dated in the badge at the top of the page. They are estimates: providers change rates frequently, and the table's headline figures exclude batch discounts, cached-input rates and negotiated enterprise pricing (batch and caching are modeled separately by the controls in step 1).

Long-context tiers are modeled. Some providers — Google's Gemini Pro models today — bill roughly double once a single prompt passes a threshold (200K tokens for Gemini). Those rows carry a tiered marker, and the moment your context requirement in step 2 crosses the threshold the table, the finder and every projection switch to the higher published rate rather than quietly quoting the cheap one. The Anthropic 1M-context models are billed at their standard rate across the whole window, so they carry no tier.

The Anthropic and Google rows were re-checked against provider pricing pages on 26 August 2026; the rest were compiled in June 2026 and are older. If you spot a stale number, use the "suggest a correction" link under the table.

Value = our editorial quality estimate (0–100, weighted heavily) divided by your per-request cost at the current slider settings, normalized so the best model in view scores 100. It rewards capable-but-cheap models and updates live as you move the sliders — a model that wins at 400 output tokens may lose at 4,000. Quality estimates are editorial judgments blending public benchmarks and community evals, not an official benchmark.

Generating a token requires a full forward pass through the model, sequentially, one token at a time — while input tokens are processed in parallel in a single pass. Output is therefore far more compute-intensive per token, and pricing reflects that. Practical upshot: setting a sensible max_tokens and asking for concise answers is one of the highest-leverage cost optimizations available.

Add up: system prompt + conversation history you keep + retrieved documents + the user's message + the maximum response you allow. For a typical chatbot that's 4K–32K tokens; for RAG over long documents, 64K–200K; for whole-codebase or book-length work, 500K+. Note that many models get slower and slightly less accurate near their context limit, so "fits" isn't the same as "optimal". Some providers also charge a premium rate above a prompt-size threshold; where that rate is published, the finder above applies it automatically and flags it on the card, so the cheapest-fit figure is the rate you would actually be billed.

Five proven levers: (1) trim your system prompt — at 1,000 requests/day, every 100 tokens removed saves ~3M input tokens a month; (2) enable prompt caching for stable prefixes; (3) use batch APIs for anything that can wait an hour (usually 50% off); (4) cap and compress outputs — ask for bullet points, not essays; (5) add a router or cascade so cheap models handle the easy 60–80% of traffic. Observability tools (see sidebar) help you find which prompts are actually burning the budget.

No. TokenTally is a fully static page — the tokenizer, calculator, chart and CSV export all run client-side in your browser. There is no backend, no analytics on your prompt text, and nothing is transmitted when you type or paste.

Prompt caching lets you re-use a stable prefix (system prompt, retrieved documents, few-shot examples) so you are not billed full price for the same input on every call — cached input typically costs around a quarter of the normal rate. The Batch API runs non-urgent jobs asynchronously for roughly half price on both input and output. They stack: a heavily-cached batch workload can land near 25% of list price. Set the cached-input share and toggle Batch mode in step 1, and every monthly figure, the value scores and the chart update live.

Yes — that is the blended multi-model router in the Pro toolkit. Assign a percentage of traffic to each of up to three models (for example 70% to a budget model, 25% to a mid model, 5% to a frontier model) and TokenTally returns the blended monthly cost, the blended cost per request, and the saving versus sending 100% of traffic to your most capable model. It is the fastest way to size the payoff of a cascade or router before you build one.