← Voltar aos guias

What Are AI Tokens? A Practical Counting Guide for 2026

AI tokens are the units a model uses to process input and generate output. They affect API costs, context limits, and how much information you can send in one request. This 2026 guide explains how to count them accurately, read usage reports, and avoid common budgeting mistakes.

What is a token?

A token can represent a word, part of a word, punctuation, whitespace, or a sequence of bytes. A tokenizer converts your text into token IDs that the model processes. Token boundaries depend on the model’s tokenizer, so the same sentence can produce different counts across models.

For ordinary English prose, one token is roughly four characters, or about three-quarters of a word. A 1,000-word article might therefore contain around 1,300 tokens. These are planning estimates: code, URLs, identifiers, emoji, and different languages can have substantially different ratios.

When comparing Claude Sonnet 5.5 and GPT-6 Astra, count the actual content for each selected model. A word count alone cannot establish an exact token count.

What counts toward API token usage?

A request usually has input tokens and output tokens, but the visible message text is only part of the picture:

• Input tokens: Instructions, conversation history, the current question, retrieved documents, and other content supplied to the model.
• Request structure: Message boundaries, roles, tool definitions, and schemas can add tokens beyond your visible prose.
• Output tokens: Generated responses, including tool-call arguments where applicable.
• Reasoning tokens: Some models report internal reasoning usage. Whether these tokens are included in output totals, billed, or exposed separately depends on the model and API.
• Multimodal input: Images, audio, and video use model-specific accounting. Text character estimates do not apply.

Repeated conversation history generally counts again when you send it in another request, unless the provider applies eligible caching or another documented mechanism.

How to count tokens accurately

• Select the exact model and endpoint. Tokenizers and request formats can differ, even when both endpoints accept similar message objects.
• Build the complete request. Include system instructions, previous messages, tool schemas, attachments, and any documents your application inserts automatically.
• Use a supported counting method. Prefer the provider’s token-counting endpoint when available. Otherwise, use a tokenizer that explicitly supports the selected model. Local text counts may exclude message-format overhead.
• Inspect a real response. In many OpenAI-compatible responses, usage appears under prompt_tokens, completion_tokens, and total_tokens. Other APIs use input_tokens and output_tokens. Check the endpoint’s documented fields.
• Log usage per request. Record the model, usage breakdown, applicable rates, and request identifier. For streaming, confirm whether usage requires an option or arrives in a final event.

Use the returned usage report to reconcile actual consumption. Compare it with billing records when you need to verify charges; a local estimate is useful for planning, but may omit provider-specific accounting.

How tokens translate into API costs

For a request without caching or special charges, cost equals input tokens divided by 1,000,000 times the input rate, plus output tokens divided by 1,000,000 times the output rate.

For example, assume hypothetical prices of $1 per million input tokens and $4 per million output tokens. A request using 10,000 input tokens and 2,000 output tokens costs $0.018: $0.010 for input plus $0.008 for output. These are illustrative rates, not a 59API price quote.

Cached reads and cache writes may have separate prices. Check whether cached tokens are already included in the reported input total before calculating charges. Likewise, do not add reasoning tokens twice if they are already included in output usage.

Reduce usage without losing useful context

Set a sensible output limit, retrieve only relevant document passages, remove duplicated instructions, and summarize older conversation turns when full detail is unnecessary. Keep reusable prompt sections stable where caching is supported. Leave room for generated output within the model’s documented context limits, including reasoning where applicable.

59API provides low-cost, pay-as-you-go access to multiple model families through https://api.59api.com, with compatibility for OpenAI SDKs, Claude Code, Codex, and Gemini CLI. It offers native, official-quality models without downgrading and a referral rebate. Sign up for 59API and compare the same representative workload across suitable models, measuring answer quality and actual cost per completed task.

Preços · Docs