Guide · Last verified 2026-08-08

How Much Does 1 Million AI Tokens Cost?

Short answer: $0.30–$5.00 per million input tokens and $2.50–$30.00 per million output tokens across the major hosted models. Output is where the money goes — here is the full table, a live calculator, and how to pay less.

Quick answer

1M token cost by model

Standard on-demand text rates per 1,000,000 tokens. Caching, batch and long-context rules can change the real bill — check the notes and the linked official sources.

Model1M input tokens1M output tokensContext windowSource
GPT-5.6 Sol
OpenAI
$5.00$30.001,050,000OpenAI
GPT-5.6 Terra
OpenAI
$2.50$15.001,050,000OpenAI
GPT-5.6 Luna
OpenAI
$1.00$6.001,050,000OpenAI
Claude Opus 5
Anthropic
$5.00$25.001,000,000Anthropic
Claude Sonnet 5
Anthropic
$2.00 *$10.00 *1,000,000Anthropic
Claude Haiku 4.5
Anthropic
$1.00$5.00200,000Anthropic
Gemini 3.6 Flash
Google
$1.50$7.501,048,576Google
Gemini 3.5 Flash-LiteLowest
Google
$0.30$2.501,048,576Google

* Claude Sonnet 5 introductory pricing of $2.00/$10.00 runs through Aug 31, 2026, then moves to $3.00/$15.00. OpenAI GPT-5.6 models bill higher long-context rates on prompts over 272K input tokens. Prices are informational estimates in USD; the provider invoice is the source of truth.

Live calculator

Price any token volume

Prefilled with 1,000,000 input tokens. Adjust input and output volumes to model your real workload.

Cheapest total
Most expensive total
ModelInput costOutput costTotal
Standard rates: text token pricing. Excluded: caching, batch, tools and taxes. Pricing data verified —
Input vs output

Why 1M output tokens cost 3–8x more than 1M input tokens.

Reading your prompt is cheap; writing the answer is expensive. Input tokens are processed in parallel in one pass, while output tokens are generated one at a time, each requiring a full model forward pass. Providers price that asymmetry directly into the rate card.

Practical takeaway: workloads that are output-heavy (long-form writing, code generation, detailed analysis) cost far more than input-heavy workloads (classification, extraction, short answers) at the same total token count. When you estimate a budget, always split input and output — never price "1M tokens" as a single blended number.

Input-heavy exampleModerating 1M tokens of user comments with 50-token verdicts: input dominates the bill.
Output-heavy exampleTurning 100K tokens of notes into 900K tokens of documentation: output is ~90% of the cost.
Budget ruleModel your input:output ratio first, then pick the cheapest model at that ratio with the API Cost Calculator.
Scale check

What does 1 million tokens actually look like?

For English text, one token averages roughly 0.75 words, so 1M tokens is about 750,000 words — around 1,500 book pages, 3–4 full novels, or roughly 4 MB of plain text. Chinese, Japanese and Korean text tokenizes denser, often near one token per character, so 1M tokens covers closer to 950,000 CJK characters.

In workload terms: a customer-support bot exchanging ~2,000 tokens per conversation burns 1M tokens every 500 conversations. A coding assistant sending a 50K-token repository snapshot per request spends 1M tokens in just 20 requests. Use the Tokens ↔ Words converter for exact planning numbers.

≈ 750,000 wordsEnglish prose at the common 0.75 words-per-token average.
≈ 1,500 pagesAt ~500 words per page. Try the converter for your own page size.
≈ 500 chat sessionsAt ~2,000 total tokens per support conversation.
Pay less

Five ways to cut the cost of your next 1M tokens

The headline rate is rarely what heavy users actually pay. These levers routinely cut real bills by 30–80%.

1. Route to smaller modelsFlash-Lite / Haiku-class models handle classification, extraction and short answers at $0.30–$1.00 per 1M input — up to 16x cheaper than flagship tiers. Reserve flagships for hard reasoning.
2. Use prompt cachingRepeated system prompts, documents and few-shot examples are billed as cached input at a steep discount on most providers. If the same prefix appears in every request, caching is usually the single biggest saving.
3. Batch non-urgent jobsBatch APIs trade latency (hours) for price — commonly around 50% off. Nightly summarization, embedding prep and evaluation runs are ideal candidates.
4. Trim the prompt itselfRedundant instructions, verbose JSON scaffolding and raw HTML waste tokens on every call. Stripping markup to clean Markdown alone often saves 30–60% on web-derived content. Measure your own text with the Token Counter.
5. Cap the outputOutput is the expensive direction. Set explicit max-output limits and ask for structured, bounded answers instead of open-ended essays.
FAQ

1M token cost questions

How much does 1 million input tokens cost?

Between $0.30 and $5.00 at standard rates among the models tracked on this page — from Gemini 3.5 Flash-Lite at $0.30 up to GPT-5.6 Sol and Claude Opus 5 at $5.00 per million input tokens.

How much does 1 million output tokens cost?

Between $2.50 and $30.00 per million output tokens. Output is billed 3–8x higher than input because generation is the expensive part of inference.

Is 1 million tokens a lot?

For English text it is roughly 750,000 words — about 1,500 pages. For a chatbot it equals hundreds to thousands of conversations; for a coding tool sending large repository context, it can disappear in a few dozen requests.

Do cached or batch tokens cost less?

Yes. Most providers discount repeated prompt prefixes (prompt caching) and offline batch jobs, often by 50% or more. This page shows standard on-demand rates, so heavy workloads usually cost less than the headline number.

Will these prices change?

Yes — LLM pricing has consistently moved down over time, and providers adjust rates and model lineups regularly. We re-verify against official pricing pages weekly; the "Last verified" date at the top shows the latest check.

Methodology

Where these numbers come from.

All rates are taken from official provider pricing pages and re-verified weekly. The calculator multiplies your token volumes by the published per-million-token rates; it does not estimate tokenization itself and does not apply caching, batch or tiered discounts unless noted.

Official sources: OpenAI model docs, Anthropic pricing, Google Gemini API pricing. Always confirm the provider invoice before making purchasing decisions.

Next step: budget itTurn this rate card into a monthly estimate with the API Cost Calculator.
Next step: measure itPaste your real prompt into the AI Token Counter.
Next step: fit itCheck whether your document fits each model with the Context Window Checker.