LLM Cost Calculator

LLM API pricing is billed per token, with separate rates for input and output. This calculator models your real usage pattern: daily requests, average tokens per call, and days per month. It then applies published June 2026 rates for GPT-4o, GPT-4.1, Claude, Gemini, DeepSeek, and Mistral so you can compare monthly and annual spend before you commit to a provider.

Use it to answer common budget questions: how much will my AI app cost, which model is cheapest at my volume, and what is the cost per request or per user. The comparison table sorts all models by monthly cost and highlights the cheapest option automatically.

LLM pricing is the per-token rate providers charge when you call their API. Unlike flat SaaS subscriptions, API billing scales with usage. Every prompt you send and every response you receive consumes tokens, and those tokens are multiplied by model-specific rates.

Flagship models like GPT-4o and Claude Opus cost more because they use larger architectures and deliver higher quality on complex tasks. Nano, mini, and flash tiers trade some capability for dramatically lower per-token rates, which makes them ideal for classification, extraction, and high-volume chatbots.

Providers tokenize your text into subword units before processing. You are charged for every token in your prompt (input) and every token the model generates (output). System prompts, conversation history, and retrieved context all count toward input tokens.

Billing is calculated as (input tokens / 1,000,000) x input price plus (output tokens / 1,000,000) x output price. There is no per-request flat fee on standard API tiers. A single long request with a large context window can cost more than many short requests combined.

Input tokens are everything you send to the model: user messages, system instructions, tool results, and RAG context. Output tokens are the model completion. Output is priced higher on nearly every model because generation requires sequential compute for each token produced.

Apps that generate long responses (summaries, code, reports) skew toward output cost. Apps that send large context but expect short answers (classification, routing) skew toward input cost. The breakdown chart in this calculator shows which side drives your bill for the selected model.

Start with realistic averages: count tokens in a typical request using the AI token counter, multiply by expected daily requests, and multiply again by working days per month. Add 10 to 20% buffer for retries, tool calls, and context growth.

Run the estimate on both your primary model and the cheapest model in the comparison table. The gap between them is your potential savings from model tiering. For multi-model apps, estimate each endpoint separately and sum the results.

The fastest savings come from model selection: move routing, classification, and FAQ tasks to GPT-4.1 Nano or Gemini Flash. Cap max_tokens on output-heavy endpoints. Cache identical system prompts where your provider supports it.

Trim RAG context to only the chunks you need. Batch overnight jobs through provider batch APIs where available for 50% discounts. Monitor cost per request weekly and set alerts before you hit budget limits.

At typical chat volumes, Gemini Flash is often the lowest monthly cost thanks to $0.075 input and $0.30 output per million tokens. GPT-4o sits in the mid tier at $2.50 / $10.00 and offers strong general performance. Claude Sonnet at $3.00 / $15.00 costs slightly more per request but excels at coding and long-context work with 200K tokens.

DeepSeek Chat and Reasoner undercut all three on price for developers comfortable with their data policies. For a detailed provider breakdown, see the LLM cost comparison guide. Provider-specific calculators: OpenAI, Claude.

Set a monthly API budget before launch and track cost per active user from day one. Separate development and production keys so test traffic does not inflate forecasts. Re-run this calculator when you change models, add features, or onboard new user segments.

Plan for 2x to 3x growth in the first quarter if your product gains traction. Use the advanced growth field to stress-test higher volume. Keep a fallback model configured so you can downgrade quickly if costs spike unexpectedly.