Claude API pricing: Opus, Sonnet and Fable per 1M tokens

Claude API prices are per million tokens, and on each of the three newest models output costs five times input. Sonnet is the cheapest of them, Opus costs twice as much, and Fable five times as much. Caching, batch and fast mode then change the bill. Below: the official rates, how a bill adds up, and our lower prices.

What does the Claude API cost per million tokens?#

Anthropic prices every model per million tokens, with separate rates for input, output and cache traffic. These are the list prices for claude-opus-5.5, claude-sonnet-5.5 and claude-fable-5.1, with the modifiers that apply on top.

Three things on that page are easy to miss. First, cache reads on the newest models cost less than the usual tenth of the input price: a twentieth on Opus and Sonnet, and a fortieth on Fable. Second, the full 1M-token context window is billed at the standard rate, so a very long prompt costs more only because it carries more tokens, not because each token costs more. Third, fast mode on Opus is a separate, higher price for faster output, and it cannot be combined with the Batch API.

How is a Claude API bill calculated?#

Each request is the sum of up to five parts, each with its own rate:

• Input tokens: everything you send that is not cached, including the system prompt, the conversation so far, and the tool definitions in the tools parameter.
• Output tokens: everything the model writes back, and thinking counts here. The response field usage.output_tokens_details.thinking_tokens shows how many of the billed output tokens were reasoning.
• Cache writes: the first time you store a prompt prefix, it costs 1.25 times the input rate for a five-minute cache, or twice the input rate for a one-hour cache.
• Cache reads: later requests that reuse the stored prefix pay the much lower cache-read rate instead of the input rate.
• Server tools: web search is charged per search, on top of the tokens its results add. Code execution adds no charge when used with the current web search or web fetch tools; otherwise it is billed by execution time.

Then the modifiers stack. The Batch API halves input and output prices for requests that can wait. US-only inference through inference_geo multiplies every token category by 1.1. Caching multipliers apply on top of both.

One detail matters for long agent sessions. On recent models, thinking blocks from earlier turns stay in the context and are billed as input on the next turn. A long chat with heavy reasoning grows its input side faster than you might expect, which is where caching pays back. Thinking on these models is adaptive, and the effort setting controls how deep it goes: lower effort means fewer thinking tokens billed as output, and the model may skip thinking on easy requests.

A worked caching example#

Say an agent sends the same 100,000-token codebase summary ten times within five minutes on Sonnet. Without caching, you pay for 1,000,000 input tokens at the full input rate. With a five-minute cache, you pay for one write at 1.25 times the input rate and nine reads at a twentieth of it. In input-rate units that is 125,000 plus 45,000, or 170,000 instead of 1,000,000: about 83% less on that part of the bill. Output does not change, so your real saving depends on how much the model writes. Prompt caching shows how to turn it on.

What do the same models cost through 5abeer?#

5abeer sells the same Claude models through an Anthropic-compatible API, paid from your balance, below the official list price. The blocks below read our current price and the official price from the live catalog, so they stay correct when either one changes.

The Anthropic models we offer on the API today:

A typical month for a small coding agent, with 20 million input tokens and 2 million output tokens on Sonnet:

That example leaves caching out on purpose. If much of your input repeats, our cache rates are listed per model on the Claude API page, and the caching guide shows how to mark the prefix you want stored.

Which Claude model should you pick?#

Price per token is only half the decision. A cheaper model that needs three attempts costs more than a stronger one that gets it right the first time. Anthropic's own advice is Sonnet for most production workloads, Opus for the most complex reasoning, and Haiku for simple tasks. For the three newest models, that works out like this:

• claude-sonnet-5.5 is the default for most work: coding agents, chat products, extraction and summaries. It is the cheapest of the three.
• claude-opus-5.5 is for the hardest reasoning, large refactors and long agent runs where a wrong step is expensive. At twice Sonnet's token price, use it where you have measured a quality gap. It is also the only one of the three with fast mode.
• claude-fable-5.1 is the top tier, at five times Sonnet's token price. Its cache reads cost only two and a half times Sonnet's, so the gap narrows for long, stable prompts that are read many times, such as a large reference corpus.

A common pattern is to route by task: Sonnet for the bulk of traffic, with Opus or Fable called only for the steps that need them. Because every model is billed per token from the same balance, switching costs nothing beyond the tokens.

How do you buy Claude API access through 5abeer?#

1. Create a key in your dashboard. API keys explains the spend limit, expiry date and IP allow-list you can set on each key.
2. Add balance. API usage is deducted from your balance, token by token.
3. Point your client at our Anthropic-compatible base URL and use the model ids above. The Claude API page lists the base URL and current prices.

The batch discount above is Anthropic's. Our docs do not list a batch endpoint, so plan your costs with us on regular requests. To compare Claude with the OpenAI and Gemini models on the same balance, see the API page. For the app plans, see pricing.

Questions#

Are thinking tokens billed as output?#

Yes. Thinking is part of the billed output tokens, and the usage field in each response shows how many were reasoning. On later turns, earlier thinking blocks stay in the context and are billed as input.

Does a long prompt cost more per token?#

Not on these three models. Anthropic states that a 900k-token request is billed at the same per-token rate as a 9k-token request. A long prompt costs more only because it has more tokens.

Is the Batch API worth it?#

If your requests can wait, it halves input and output prices at Anthropic, and it combines with caching. It does not work with fast mode. Our docs do not list a batch endpoint.

How is web search charged?#

Each search has a fixed price, shown in the table above, plus the tokens its results add to the context. A search that fails with an error is not billed.

How much do cache reads cost on the newest models?#

A twentieth of the input price on the newest Opus and Sonnet, and a fortieth on the newest Fable, against the usual tenth on other models. A cache write costs more than a normal input token, so caching pays off only when a prefix is read again.

Is a subscription cheaper than the API?#

It depends on how much you use and where. Our Claude plans, Max (formerly Max 5x) and Ultra (formerly Max 20x), work inside the apps, while the API is paid separately from your balance. Claude API vs subscription and the cheapest way to run Claude Code compare them with worked numbers.

Sources#

Checked on 2026-11-01:

• Pricing, Claude API docs
• Extended thinking, Claude API docs