Claude API pricing, explained

The Claude API is pay per token with no subscription and no monthly minimum: you are billed separately for input and output, quoted per million tokens, on exactly what you send and receive. As of August 2026 input runs from $1 per million tokens on Claude Haiku 4.5 to $10 on Claude Fable 5, and output is priced at five times input on every current model. Here is the full rate card, the discounts that actually change the math, and three worked examples you can substitute your own numbers into.

What the Claude API costs per model

Anthropic prices the Claude Developer Platform per million tokens (MTok), quoted as input / output. As of August 2026:

• Claude Opus 5: $5 in / $25 out per MTok
• Claude Sonnet 5: $2 in / $10 out per MTok
• Claude Haiku 4.5: $1 in / $5 out per MTok
• Claude Fable 5: $10 in / $50 out per MTok
• Claude Opus 4.8, 4.7, 4.6 and 4.5: $5 in / $25 out per MTok
• Claude Sonnet 4.6 and 4.5: $3 in / $15 out per MTok

Sonnet 5 launched at $2 / $10 as introductory pricing through August 31, 2026; Anthropic confirmed on August 10, 2026 that this is now the standard price and the scheduled increase to $3 / $15 will not happen. The 1M-token context window on Claude 4.6 and later is billed at standard per-token rates, so a 900k-token request costs the same per token as a 9k-token one; there is no long-context premium tier on the current models.

Two things to get right when you pick a row. Newest is not the same as most capable: Opus 5 is the most recent release, while Claude Fable 5 is Anthropic's most capable widely released model, which is why it sits at the top of the rate card. And several older models are retired, so do not budget against their old rates: Opus 4.1 retired on August 5, 2026, Opus 4 and Sonnet 4 on June 15, 2026, and Claude 3.7 Sonnet and Claude 3.5 Haiku on February 19, 2026.

These numbers move, so treat the list above as a snapshot and check platform.claude.com/docs/en/about-claude/pricing before you put a figure in a budget. Most accounts pay with prepaid usage credits that you buy in the Console and that draw down as you use the API; organizations with an invoicing arrangement are billed monthly instead. New accounts get a small amount of free credits to test with, and there is nothing to subscribe to. If you assumed your Claude Pro plan covered this, it does not: see does Claude Pro include API credits.

Input versus output: output is what costs you

Output costs five times input on every current model. Opus 5 is $5 in and $25 out, Sonnet 5 is $2 and $10, Haiku 4.5 is $1 and $5. The ratio holds across the range, which means response length drives cost more than prompt length does. A 20,000-token prompt that returns one paragraph is cheap. A 2,000-token prompt that returns a 5,000-token document is not. If you want a smaller invoice, the first lever is usually asking for shorter answers, not writing shorter prompts.

One token is about four characters, or roughly 0.75 English words. So a thousand words of English is around 1,300 tokens, and a million input tokens is on the order of 750,000 words. Use that to sanity-check an estimate before building a spreadsheet around it.

The discounts that actually change the math

Three mechanisms move the number enough to matter, and all three are opt-in. Rates below are as of August 2026.

• Prompt caching. Multipliers apply to the model's base input rate: a five-minute cache write costs 1.25x, a one-hour cache write costs 2x, and a cache read costs 0.1x. A five-minute cache pays for itself after a single read: you spend 0.25x extra to write it and save 0.9x on the first hit. For anything with a stable system prompt, a long document, or a repeated codebase context, it is the largest single lever available.
• The Batch API. 50% off both input and output in exchange for asynchronous processing. If the work does not have to answer a human in real time (classification, enrichment, evals, bulk summarization), batching halves the bill with no change to the output.
• Server tools. Web search is billed at $10 per 1,000 searches on top of tokens. Web fetch adds no charge. Code execution is free when used alongside web search or web fetch; on its own, each organization gets 1,550 free hours per month, then $0.05 per hour per container.

Caching and batching stack. A batched workload with a cached prefix is the cheapest way to run the API at volume.

Three worked monthly examples

The volumes below are illustrative, not measured. The point is the arithmetic: swap in your own token counts. All three use August 2026 rates and ignore caching unless stated.

1. Light chatbot workload on Haiku 4.5 ($1 / $5). Say 2,000 conversations a month at about 1,500 input tokens and 500 output tokens each.

• Input: 2,000 conversations x 1,500 tokens = 3,000,000 tokens = 3 MTok x $1 = $3
• Output: 2,000 x 500 = 1,000,000 tokens = 1 MTok x $5 = $5
• Monthly total: about $8

2. Document processing on Sonnet 5 ($2 / $10). Say 5,000 documents a month at about 8,000 input tokens in and 1,200 output tokens of structured summary out.

• Input: 5,000 x 8,000 = 40,000,000 tokens = 40 MTok x $2 = $80
• Output: 5,000 x 1,200 = 6,000,000 tokens = 6 MTok x $10 = $60
• Monthly total: about $140, or roughly $70 if the whole job runs through the Batch API at 50% off

3. Heavy agentic coding on Opus 5 ($5 / $25). Say 500 sessions a month at about 40 turns each; because an agent resends its accumulated context every turn, call it 30,000 input tokens and 800 output tokens per turn.

• Input: 500 x 40 x 30,000 = 600,000,000 tokens = 600 MTok x $5 = $3,000
• Output: 500 x 40 x 800 = 16,000,000 tokens = 16 MTok x $25 = $400
• Monthly total: about $3,400

Now apply caching to the third example. If 80% of that input arrives as cache reads at 0.1x, the input line becomes 120 MTok at $5 ($600) plus 480 MTok at $0.50 ($240), so roughly $840 instead of $3,000, before the modest write premium. The total lands near $1,240, which is why caching is the first thing to reach for on agentic work.

Why is the Claude API so expensive?

The sticker shock is real, and the reasons are structural.

• Output dominates. At five times the input rate, a verbose model is an expensive model. Bills blow up on generation, not on reading.
• Agentic loops resend context every turn. A 40-turn session does not send your prompt once; it sends a growing transcript 40 times. Input volume compounds with conversation length unless you cache or prune, which is why coding agents cost far more than chat at the same rate.
• The tokenizer changed. Claude 4.7 and later use a newer tokenizer that produces roughly 30% more tokens for the same text. A per-token price comparison across tokenizer generations therefore misleads: the same page of text is counted differently, so cost per page moved even where the posted rate did not.

The levers, in the order worth trying: cache the stable prefix, batch anything asynchronous, route routine work to a smaller model (Haiku 4.5 at $1 / $5 handles plenty of what people send to Opus by reflex), and cap output length in the prompt. One more people skip: if a subscription you already pay for covers the work, that surplus is cheaper than any API rate, because it is already spent. The comparison is in Claude Pro vs API: the real cost.

The cheapest tokens are the ones you already bought

Metered billing and a flat subscription fail in opposite directions. On the API an idle day costs nothing. On Claude Pro or Claude Max an idle day is paid-for capacity you never get back: subscription limits are a ceiling, not a balance, and whatever you do not use before the window resets is gone.

That recurring surplus is what Tokens for Good puts to work. Your Claude Code claims a queued nonprofit, researches its real-world impact against a fixed methodology with citations, and submits a structured report. Every organization is researched twice by independent contributors, validated, consolidated, scored deterministically, and human-reviewed before it reaches the public directory. It runs on the subscription you already pay for with no separate API cost, and it can run on a schedule in the background. "Tokens" here means AI model tokens, not crypto: no coin, no wallet, no blockchain. For more ways to spend the leftovers, see what to do with unused Claude capacity.

Frequently asked questions