Input Tokens vs Output Tokens: What Counts and What Costs

Input and output token comparison

Neither side is automatically the largest cost. A document-review request can be input-heavy, while a short prompt that asks for a long report can be output-heavy. Measure the request shape that your product actually sends.

Use the Token Calculator to measure prompt size, then use the AI API Cost Calculator to combine input tokens, output tokens, model prices, request volume, and active users.

How much more do output tokens cost?

Across 74 current text models with published input and output prices, output tokens cost a median 5.0× the input price (range 1.0×–10.0×); output is more expensive for 96% of them, with prices verified Jul 7, 2026 to Nov 6, 2026.

Real request-cost example from current pricing data

The example below uses a current model from AICostBudget's source-linked pricing data and recalculates whenever that data is updated. The model name, input price, output price, cached-input price when available, and arithmetic come from the data, not from hand-typed numbers.

What counts as input tokens?

• The current user prompt or message.
• System and developer instructions included in the API request.
• Earlier conversation turns that are sent again with the current request.
• Retrieved documents, help-center passages, database results, or other RAG context.
• Few-shot examples, tool definitions, schemas, and formatting instructions sent to the model.

A short visible user message can still create a large input bill when the application attaches a long system prompt, full conversation history, several retrieved documents, and tool schemas. Inspect the complete API payload, not only the text box the user sees.

How many tokens is an Excel or CSV file?

A model can't read an .xlsx file as it sits on disk. Before it sees anything, the cells get turned into text, either by your own code or by a provider's file feature. You pay input tokens for that text, not for the file size, so the format the rows end up in decides most of the bill.

To put numbers on it, we built a 1,000-row, 9-column order export (order ID, date, city, product, quantity, unit price, total, status, and a short note on some rows) and counted it with OpenAI's open o200k_base tokenizer. The data is synthetic, and Claude and Gemini use their own tokenizers, so treat the counts as a guide rather than an exact bill.

CSV and TSV come out about the same. A Markdown table adds 14% for the pipes and spaces. JSON is the expensive choice because every row repeats every column name: minified JSON costs 1.76 times as much as CSV, and pretty-printed JSON 2.62 times as much, for exactly the same data.

Format shows up in the monthly bill too. Sending that export once a day for 30 days and asking for a 1,000-token summary each time costs about $2.42 on Claude Sonnet 5.5 as CSV and $5.84 as pretty-printed JSON, using the token counts above. To count your own file row by row, upload it to the CSV Token Counter. Its estimate counts digits at their own measured rate, so it does not need the 1.7× correction.

• Send only the columns the task needs. IDs, timestamps, and internal codes are mostly digits, and digits are the most token-hungry text in the sheet.
• Send CSV or TSV, not JSON, unless the model has to return structured records.
• Aggregate before sending: a pivot table or a few summary rows often answers the question that 1,000 raw rows would.
• For classification or extraction, send rows in batches and reuse one short instruction, instead of repeating the instructions on every row.

What counts as output tokens?

Output tokens are the response content generated by the model and returned under the provider's billing rules. This can include natural-language answers, summaries, code, classifications, or structured output such as JSON.

Provider treatment of internal or reasoning-related tokens can differ by model and API. Do not assume hidden-token behavior is identical across providers; verify the documentation and billing records for the exact model you use.

Why output tokens can matter more without always costing more

Many models list different input and output prices, and some list a higher output rate. That makes generated length an important cost lever, especially for reports, code, long explanations, and agent responses. It does not mean output is universally more expensive or that output always dominates the bill.

Compare the exact price structure for candidate models on the Model Pricing Comparison page. Keep the same token assumptions when comparing models so a cheaper-looking rate is not offset by longer responses, retries, or a different workflow.

How cached input changes the calculation

Some models in the canonical pricing data expose a cached-input price. Cached input represents eligible repeated request content billed at the provider's listed cached rate instead of the normal input rate. Stable system instructions, repeated document prefixes, or other reusable context may qualify, depending on the provider and model.

Request cost with eligible cached input =
(regular input tokens / 1,000,000 x regular input price)
+ (cached input tokens / 1,000,000 x cached input price)
+ (output tokens / 1,000,000 x output price)

How to reduce input and output token cost

• Remove duplicated instructions and irrelevant retrieved context.
• Summarize or window old conversation history instead of resending everything.
• Measure representative prompts before changing production behavior.
• Constrain output length when a concise answer satisfies the product requirement.
• Use cached input only where the selected provider and model support it.
• Compare input, cached-input, and output price structures before changing model routes.
• Track retries and fallback calls because they can repeat both input and output cost.

First measure sample text in the Token Calculator, then calculate the full request and monthly usage in the AI API Cost Calculator. Token budgeting still matters even when an offer claims access is free; review the risks behind free or unlimited OpenAI API key claims before sending real product traffic.

FAQ