Anthropic API Pricing 2026: Claude Model Costs, Token Billing, and How to Save

Anthropic API Pricing 2026: Claude Model Costs, Token Billing, and How to Save

Every price figure in this article is taken from Anthropic's official pricing page (platform.claude.com/docs/en/about-claude/pricing), as of 2026-10-09. Prices and model lifecycles change; before you build a budget, check the official page for the current figures.

Before you call the Claude API, get two things straight: what the unit price is, and how the bill is calculated. This article works through the official price list of the Claude models on sale in 2026 — input/output unit prices, the token billing basis, and four officially documented ways to save. All prices are in USD per million tokens (MTok).

A quick conclusion first: if you just want to get a prototype running, starting from the Haiku or Sonnet tier is enough; what really determines the size of the bill is usually not the price table, but how much context you carry into each request and how long you let the model output. Below we go step by step through "price → billing → savings → getting started".

1. Price overview: unit prices for mainstream Claude models (as of 2026-10-09)

The table below shows the Base input / Output unit prices from Anthropic's official pricing page, omitting the cache columns (see section 3):

Two official points that are easy to trip over:

• Previous-generation models: Opus 4.1, Opus 4, Sonnet 4, Haiku 3.5 and the like are the prior generation; the official pricing page still lists their historical unit prices (e.g. Opus 4.1 $15/$75, Sonnet 4 $3/$15, Haiku 3.5 $0.80/$4), and some channels no longer sell them, keeping them only on Bedrock / Google Cloud and similar. For new projects, pick from the on-sale models in the table above; if you really need a previous-generation model, first check the official model lifecycle page to confirm availability and the current price.

• New tokenizer: Anthropic notes that Claude 4.7 and later models use a new tokenizer that produces about 30% more tokens for the same text (the actual ratio varies with the content). When comparing prices across generations, factor this in — otherwise "the unit price dropped but the bill didn't".

2. How token billing works

Pay-as-you-go, with input and output billed separately. The cost of one request = input tokens × input unit price + output tokens × output unit price. The official FAQ's rough conversion: 1 token is about 4 English characters or 0.75 English words; the ratio differs for Chinese and other languages, so rely on the actual tokenizer count.

Looking at the current price list, across the whole lineup the output unit price is 5× the input price (e.g. Sonnet 5.5 $2 → $10). This means: the longer the output, the steeper the bill. When writing prompts, cutting filler and asking for concise output is itself a way to save.

A concrete example. Suppose you run customer-support Q&A with Claude Sonnet 5.5: each request carries 10K tokens of system prompt and knowledge-base snippets (input), and the model outputs 2K tokens. Cost per call = 10,000 × $2 / 1M + 2,000 × $10 / 1M = $0.02 + $0.02 = $0.04. A day of 10,000 such requests is $400. Note that input and output each account for exactly half of the total — if you cut the output to 500 tokens, the per-call cost immediately drops to $0.025, and the monthly savings are considerable. When budgeting, first estimate the "input/output volume per request" and multiply by the unit price; that is far more reliable than staring at the price table alone.

On long context, the official position is: for Claude 4.6 and later models (except Haiku 5.5), the 1M-token context window is billed at the standard unit price, with no long-context surcharge — a 900K-token request has the same unit price as a 9K-token request. Haiku 5.5 is the exception: it is tiered by prompt length, and a prompt over 100K tokens (including cache reads/writes) is billed entirely at the higher tier.

3. Four ways to save (all officially documented)

1. Prompt Caching. Write the system prompt, documents, and long conversation prefixes into the cache, and subsequent requests read them at the cache price. Official multipliers (relative to the base input price): 5-minute cache write = 1.25×, 1-hour cache write = 2×, cache-hit read = 0.1× (0.05× for Opus 5.5 / Sonnet 5.5, 0.025× for Fable 5.1 / Mythos 5.1). In Anthropic's words, the 5-minute tier pays for itself after a single cache read. Turn this on first for multi-turn agents and long-system-prompt scenarios.

2. Batch API. Asynchronous processing, with input and output uniformly 50% off (the official page states that Batch reduces both input and output by 50%). For example, the Sonnet 5.5 batch price is $1 / $5, and the Haiku 4.5 batch price is $0.50 / $2.50. Tasks that don't need results immediately — offline evaluation, bulk rewriting, data preprocessing — should all go through batch.

3. Choose the model by task. The official cost advice is exactly tiered usage: Haiku for simple tasks, Sonnet as the production workhorse, and Opus only for complex reasoning. The price gap is an order of magnitude: Haiku 5.5 (prompt ≤ 100K) input is $0.10/MTok, 1/40 of Opus 5.5 ($4). The one-step routing approach is to explicitly route "simple requests" to small models: hand classification, extraction, and formatting to small models, and let only requests that genuinely need multi-step reasoning reach the large model — the average cost of a mixed workload drops noticeably.

4. Avoid hidden surcharges. The surcharges Anthropic lists include: inference_geo: "us" (US-only inference) at ×1.1 on 4.6+ models; a Fast mode premium (Opus 5.5 $8/$40, Opus 5 and 4.8 $10/$50); the server-side web search tool at $10 / 1,000 searches. Default to global routing and enable fast mode only when needed, and your budget won't blow up for no reason.

4. FAQ

Do prices change often? Yes. Anthropic may adjust prices for the same model in batches, and previous-generation models are gradually retired and taken over by partner cloud platforms. So the figures here are a snapshot (as of 2026-10-09); leave room for movement in long-term budgets and periodically check back against the official pricing page.

Can prompt caching and the Batch discount be stacked? Yes. The official FAQ states clearly that the Batch API and prompt caching discounts can be combined — for example, a batch task with a cached prefix gets both discounts at once.

Are the tools and system prompt in the input billed? Yes. Tool definitions (the name, description, and schema in the tools parameter) and the tool_use / tool_result blocks produced by tool calls are all billed normally as input or output tokens; server-side tools (such as web search) are billed separately per call. When using an agent framework, don't overlook this "invisible input".

5. No official key? Start directly with CodeGateway

Applying for an official Anthropic API key requires an overseas payment method and account system; for many developers the barrier isn't technical but payment-related. If you don't have an official key, or don't want to bind a foreign card, you can get one directly at [CodeGateway](https://www.codegateway.dev/pricing): sign up and use it right away, top up in RMB, billed against the same set of model capabilities as the official offering, supporting the full Claude lineup plus OpenAI / Anthropic-compatible endpoints.

Before getting started, check the model catalog to confirm the models you need are on sale, then use the pricing page to estimate your monthly usage. Get one real request working with a small amount first, then scale up — that is also the most direct way to verify token billing. Once it's working well, adjust the tier according to your volume.

Price data source: [Anthropic's official pricing page](https://platform.claude.com/docs/en/about-claude/pricing), as of 2026-10-09. The official page also lists billing for partner platforms such as AWS Bedrock / Microsoft Foundry (invoiced through the cloud provider, with additional surcharges on regional endpoints); this article covers only first-party Claude API billing. This article is not a price commitment of any kind; if Anthropic changes prices, the official page prevails.