Last reviewed: 2026-10-25 · Prices from our dataset, rechecked against each provider’s official page · Price history

Every model we track charges more for the tokens it writes than for the tokens it reads, from 2× to 8.3× as much. That gap, more than the headline input price, decides what many workloads cost. This page lists the ratio for each model, shows how it changes which model is cheapest, and covers the ways to keep output under control.

The ratio for every model

5× is the most common ratio (10 of 19 models), and it is the ratio Anthropic uses across its current models. The outliers are worth knowing: xAI Grok Build 0.1 charges only 2×, while Google Gemini 3.5 Flash-Lite charges 8.3×. The blended columns give one effective price per million tokens for a given mix: (input price × input share + output price × output share). A 10:1 mix is typical of chat with history, summarisation and RAG; 1:1 is closer to rewriting or translation.

Why output is priced higher

Providers do not publish their cost breakdowns, so treat this as the general technical explanation rather than any provider’s statement: a model can process all the tokens of a prompt in parallel, but it must generate its answer one token at a time, each step depending on the last. Generation keeps the hardware busy for longer per token, and the price reflects it. What the providers do document is what counts as output. Google’s Gemini price table labels the output price “including thinking tokens”, and xAI lists reasoning tokens (“internal thinking and planning”) among the token types billed for a request. A reasoning model can therefore bill far more output than the text you see.

The mix changes which model is cheaper

Compare xAI Grok Build 0.1 ($1 in, $2 out) with Google Gemini 3.8 Flash ($0.75 in, $3.75 out) on two jobs of similar size:

• Read-heavy (100M input, 10M output a month, e.g. summarising documents): Grok Build 0.1 costs $120, Gemini 3.8 Flash $113.
• Write-heavy (10M input, 40M output, e.g. drafting long articles from short briefs): Grok Build 0.1 costs $90.00, Gemini 3.8 Flash $158.

The cheaper model swaps places between the two jobs. Sorting a price table by input price alone would have picked the wrong one for one of them. The comparison table sorts by either column; the calculator takes your real input and output counts.

Where the asymmetry bites

• Long-form generation (articles, code files, reports): output dominates, so the output price is the number to shop on.
• Reasoning and agent workloads: hidden thinking tokens and the model’s own tool calls are output, and agents produce output at every step.
• Chat: per turn, input usually dominates because the history is resent, but replies that are three times longer than needed still triple the output line.
• Classification and extraction: the cheapest case. A one-word label or a small JSON object is a handful of output tokens, so input price and caching matter most (see caching and batch discounts).

How to keep output costs down

• Set a maximum output length on every request. Every major API has a max-tokens style parameter; it caps the worst case.
• Ask for the format you need. “Answer in at most three sentences” or a fixed JSON schema produces fewer tokens than open prose.
• Choose the reasoning level deliberately. Where a model lets you switch thinking off or lower its effort, simple tasks rarely need the maximum.
• Do not regenerate what you can reuse. Ask for a diff or the changed section instead of a full rewritten document.
• Batch what can wait. Batch discounts apply to output tokens too: 50% at Anthropic, OpenAI and Google, 20% on the discounted Grok models.

Common misreadings

• “Model X is cheaper because its input price is lower.” Only if your workload is input-heavy. Check both columns.
• “Cached input makes the bill tiny.” Caching only discounts input; output stays at full price, so output-heavy apps save little from it.
• “The visible answer is the output.” Not on reasoning models: the thinking tokens are billed too, and you may not see them.

Read next