Pricing

AI API pricing explained in plain English

Last verified: September 2026

What a token is (and why the math is easier than it looks)

A token is the billing unit. As a rough rule: ~750 English words ≈ 1,000 tokens, so 1 million tokens ≈ 750,000 words — roughly 1,500 pages of text. Vendors quote prices "per 1M tokens," always split two ways:

• Input tokens — what you send the model (your prompt, your documents).
• Output tokens — what the model writes back.

Output is always priced higher — typically 3–5x input — because generating text costs the model more compute than reading it.

Real-terms math: the six-cent document summary

Take a 50-page document: roughly 25,000 input tokens. Ask for a one-page summary back: roughly 1,000 output tokens. Run it on Claude Sonnet 5 ($2 per million input, $10 per million output — Anthropic made this intro rate permanent on August 11, 2026, cancelling the planned rise to $3/$15):

• Input: 25,000 tokens × $2/M = $0.05
• Output: 1,000 tokens × $10/M = $0.01
• Total: ≈ $0.06

Six cents. That's the thing to internalize: individual API calls are measured in cents. The bill only gets serious at volume — thousands of documents, millions of chatbot conversations, or heavy coding-agent runs.

Workhorse vs frontier: the July 2026 rate cards

Prices below are per 1M tokens, input/output, from the vendors' July 2026 rate cards:

The pattern: a frontier model (Claude Fable 5 at $10/$50) costs roughly 10x the input and 10x the output of its workhorse sibling (Haiku 4.5 at $1/$5). The honest question for most automated jobs is whether the frontier model's answers are 10x better for your specific task — usually, they aren't. Start on a workhorse, escalate only the calls that need it.

API billing is in USD worldwide; your card provider handles conversion. Also worth knowing: models from Claude 4.6 onward include the full 1M-token context window at standard rates — no long-context surcharge.

Batch and caching: the two big discounts

• Batch processing — ~50% off. If your job doesn't need an answer this second (overnight document processing, bulk classification), vendors' batch APIs run at roughly half the standard price. Same model, same quality, just slower.
• Prompt caching — up to 90% off. If you send the same long context repeatedly (a system prompt, a reference document, a codebase), cached input tokens can be discounted up to 90%. For chatbots and agents that re-read the same background material every turn, caching is often the single biggest line-item saver.

Subscription vs API: the rule of thumb

• Subscription ($20/month) wins for steady daily human use. One person chatting, writing, and researching all month would have to burn a lot of tokens to exceed $20 of API spend — and the subscription has no metering anxiety.
• API wins when usage is automated, spiky, or embedded in something: a script that summarizes incoming documents, an AI feature inside your product, a coding agent running in CI. Also the only option when a subscription's usage limits would choke your workload.

Note: since mid-2026, heavy programmatic use of Claude (Agent SDK, command-line, GitHub Actions) can draw from a separate credit pool billed at API rates on top of a subscription — so if you automate, read the current terms either way.