← Retour aux guides

Token Billing Explained: Estimate AI API Costs

Start with the units that AI APIs bill

AI models do not bill by request count alone. They bill primarily by tokens, small units of text that can be words, punctuation, code fragments, or parts of words. A short English word may be one token, while source code, JSON, URLs, and non-English text can tokenize differently. This is why two prompts with the same character count may cost different amounts.

Most model invoices have at least two usage categories: input tokens, which include your system prompt, chat history, user message, retrieved documents, and tool definitions; and output tokens, which are generated by the model. Input and output prices are usually different, and output tokens commonly cost more because generation requires additional compute.

When using a relay such as 59API, the practical goal is the same: send requests to the API, read the usage returned by the response, and multiply that usage by the published rate for the exact model. 59API provides pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models, without requiring developers to downgrade to a weaker substitute.

Use the core cost formula

For a standard text-generation request, estimate cost with this formula: input tokens divided by 1,000,000, multiplied by the input price per million tokens, plus output tokens divided by 1,000,000, multiplied by the output price per million tokens.

For example, assume a model costs $0.80 per million input tokens and $4.00 per million output tokens. A request containing 120,000 input tokens and producing 18,000 output tokens would cost $0.096 for input and $0.072 for output, or $0.168 total. These numbers are only an example; always use the current rate shown for your selected model before budgeting.

Do not estimate from the maximum output setting alone. A maximum is a ceiling, not actual usage. For a reliable forecast, measure average input and output tokens from representative production traffic, then also calculate a high-percentile case for unusually long conversations or documents.

Account for every billable token source

• System instructions: A long policy prompt is sent repeatedly unless your integration uses a supported caching mechanism.
• Conversation history: Each follow-up can resend prior messages, causing input usage to grow over time.
• RAG context: Retrieved chunks, metadata, citations, and duplicate passages can become the largest cost driver.
• Tool schemas and tool results: Function definitions and returned JSON are part of the context window and may add tokens.
• Reasoning, cache, or modality features: Some models expose separate categories for cached input, reasoning output, images, audio, or other capabilities. Check the model-specific billing rules rather than applying a text-only formula blindly.

Build a production cost estimate in five steps

• Choose the model first. Match capability to the job. Use a faster, lower-cost model for classification, extraction, routing, and simple support answers; reserve premium models for difficult reasoning, complex coding, or high-value outputs.
• Collect a sample. Run 100 to 1,000 realistic requests, including short, typical, and long sessions. Record model name, input tokens, output tokens, status, latency, and feature flags.
• Calculate average cost per request. Apply the published input and output rates to each record, then calculate the mean. Also calculate p95 cost so a few oversized requests do not surprise your budget.
• Project monthly volume. Multiply average cost per request by expected successful requests. Add a buffer for retries, testing, traffic growth, and longer-than-normal prompts.
• Set controls. Cap output tokens, trim old history, limit retrieved context, deduplicate documents, and alert when daily spend or p95 token usage rises.

Read usage instead of guessing

Your application should log the usage fields returned by every successful API response. Depending on the compatible API format, these may be named input tokens and output tokens, or prompt tokens and completion tokens. Store the values alongside the model and request type. This lets you answer operational questions such as which feature is expensive, whether a prompt change increased context size, and whether users are triggering abnormally long generations.

For OpenAI-compatible integrations, configure your client with the 59API base URL, https://api.59api.com, then select the intended supported model. Teams using Claude Code, Codex, or an OpenAI SDK can keep their familiar workflow while centralizing pay-as-you-go usage through one relay. Before launching, validate a small test batch and compare your locally recorded totals with the account dashboard.

Lower cost without lowering answer quality

The biggest savings usually come from reducing unnecessary input, not from aggressively shortening every answer. Keep system prompts focused, summarize old conversation turns, retrieve fewer but better document chunks, and use structured output only when your application needs it. Route simple tasks to economical models and escalate only ambiguous or high-stakes requests to stronger models.

59API is a strong option for cost-conscious teams because it combines low-cost, pay-as-you-go access with native official-quality model options and broad compatibility. Its referral rebate can also reduce effective spend for eligible referrals. When you are ready to measure real workload costs, sign up for 59API and run a controlled production-like sample before committing your monthly budget.

Tarifs · Docs