Anthropic API

Anthropic sells Claude through a pay-per-token API with no monthly fee. On 2026-10-31 the list is Opus 5.5 at $4 in and $20 out per million tokens, Sonnet 5.5 at $2 and $10, Haiku 5.5 at $0.10 and $0.50, and Fable 5.1 at $10 and $50. Content is not used for training without your permission. Fable 5.1 requires 30-day retention.

Facts at a glance

Prices checked 2026-10-31. For the full cache and batch tables, see Claude API pricing. This page is about the vendor as a supplier: what you sign up to, what the limits are, and what happens to your data.

Models offered and what they cost

Anthropic sells four current tiers. The numbers below are standard processing, per million tokens.

Three details change the bill more than the headline rate.

1. Haiku is priced by prompt length. The cheap rate stops at 100,000 input tokens. Above that the whole request moves to the second row. A pipeline that stuffs long documents into Haiku can pay five times what the price list first suggests.
2. Cache reads are a fraction of input. The pricing docs put a read at 0.1x the base input price on most models, 0.05x on Opus 5.5 and Sonnet 5.5, and 0.025x on Fable 5.1. A 5-minute cache write costs 1.25x, so a prompt prefix that is read once more has already paid for itself.
3. A newer tokenizer. Anthropic says its 4.7-and-later tokenizer produces about 30 percent more tokens than earlier models for the same text. Two models at the same dollar rate can therefore bill differently on identical input, and a move from an older Claude to a new one should be tested on your own prompts.

Older models stay on the list at older rates, for example Sonnet 4.5 and 4.6 at $3 and $15. Mythos 5.1 matches Fable 5.1 on price but has limited availability. It is open to organizations verified through programs such as Anthropic's Cyber Verification Program. If you want to know which model to use for which job, the Opus vs Sonnet page is the place to start, and Claude models covers the line itself.

Rate limits and tiers

Claude limits are set per organization and come in tiers: Start, Build, Scale and Custom. New organizations, or ones with little usage history, can begin in an Evaluation tier with limits below the standard figures while Anthropic builds an account history. The docs say tiers move up over time as you use the API, and you can ask for higher limits.

Every tier above Evaluation also carries a monthly spend cap.

At the Start tier the docs list, for the current models:

Two behaviours are worth knowing before launch day. First, limits are maximums, not guarantees. The docs say they exist to reduce unintended overspend and to share capacity fairly. Second, a limit stated per minute can be enforced in shorter slices. A rate of 60 requests per minute might be enforced as one request per second, so a burst can fail even when your minute average is fine.

When you reach a spend cap the API pauses until 00:00 UTC on the first of the next month. Requests return HTTP 429 with the error code enforced_spend_limit_reached, which lets you tell a spend stop from an ordinary rate limit. You can set your own limit below the tier cap on the Billing page. For a team, that is the simplest guard against a loop that calls the API forever.

Data retention and training

This is the part procurement teams read first, so we quote what the docs say rather than paraphrase loosely.

• Training. "Retained data is never used for model training without your express permission."
• Default storage. Conversation content, meaning your prompts and Claude's outputs, is not retained by default. The exception is the Covered Models, which need 30-day retention. As of 2026-10-31 those are Claude Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5.
• Zero data retention. Under a ZDR arrangement Anthropic does not store prompts or responses at rest after the response is returned. You request it from the sales team, and it is enabled per organization, so each new organization needs its own enablement.
• Not everything is eligible. Features that use /v1/messages but are marked "No" in the feature eligibility table, such as code execution, fall outside ZDR.
• Fable and ZDR. Fable 5.1 and Mythos 5.1 are not available under ZDR unless Anthropic expressly authorizes it.
• Claude Code. With API keys from a commercial organization, ZDR applies. If metrics logging is on, usage statistics are exempt and may be retained.

The Fable detail matters in practice. If your compliance rule is "no vendor storage of prompts", the top model is the one model line that will not meet it by default. Sonnet 5.5 and Opus 5.5 can sit under ZDR once the arrangement is in place. For the consumer apps and the Compliance API, the docs describe separate retention models, for example six years for the Activity Feed, so do not read the API rule as covering the whole company.

Regions and where else you can buy Claude

The inference_geo parameter sets where inference runs for a request. The values are us and global. You can set a default per workspace, and a workspace can be restricted to allowed geos. The parameter works on Claude 4.6 and later. On older models it returns a 400 error. Workspace geo, set when you create a workspace and fixed afterwards, can only be us today.

On Amazon Bedrock and on Google-managed Claude Models on Google Cloud, the cloud provider is the data processor, and the region comes from your endpoint or inference profile. For those, the cloud provider's own retention and compliance documents apply instead of Anthropic's. That is a real reason to pick a cloud route: you inherit the controls your security team already audited. See AWS Bedrock for that route.

Which agents can use it, and how

Claude Code is Anthropic's own agent and the obvious fit. With a commercial API key in the environment it bills the API rather than a subscription plan:

export ANTHROPIC_API_KEY=sk-ant-...
claude

Teams can push policy to every machine through managed settings, delivered as a managed-settings.json file, an MDM policy or server-managed settings from the claude.ai console. See Claude Code and Claude Code pricing for plans against API billing.

Other agents bring their own key. Aider takes it on the command line:

aider --model sonnet --api-key anthropic=<key>

OpenCode connects from inside the terminal interface with /connect, then you pick a Claude model with /models. See Aider and OpenCode. Other providers can also speak the Anthropic message format. DeepSeek and OpenRouter both document pointing Claude Code at their own base URL, which we cover on DeepSeek and OpenRouter.

Alternatives

Who should pick Anthropic

A worked cost example

Take a job that sends 10 million input tokens and gets 2 million output tokens back. The arithmetic is ours, from the published rates, and it ignores caching and tool charges.

Two things stand out. Output is five times the input price on every model, so a verbose answer costs more than a long prompt. And the spread between Haiku and Fable on this job is a factor of one hundred. Routing easy calls to Haiku and keeping Fable for the few that need it is the single biggest saving available, bigger than any caching trick. If your prompt prefix repeats, a cache read on Sonnet 5.5 costs $0.10 per million, so ten million cached tokens cost one dollar instead of twenty.

Moving to or from Anthropic

Because the Messages API is Anthropic's own format, a switch is not a one-line change the way it is between two OpenAI-compatible hosts. Plan for these steps.

1. Message format. Tool calls, system prompts and content blocks differ from the Chat Completions shape. A gateway such as OpenRouter hides this, at the price of a credit fee.
2. Model names in config. Agents that hard-code a model name need an edit. DeepSeek's Anthropic-format endpoint maps any model name starting with claude-opus to its own deepseek-v4-pro and other claude names to deepseek-flash, which is convenient and also easy to forget about.
3. Prompt tuning. A prompt tuned for one model line rarely transfers unchanged. Re-run your own evaluation set on the new model before moving traffic.
4. Spend limits and tiers. A new organization may start in the Evaluation tier. Do not assume your old limits carry over, and ask for higher limits before a launch.
5. Retention paperwork. ZDR is enabled per organization. If you split work across organizations, request it for each one.

If you are leaving Claude, the cache-read price is the number most teams miss. Agent loops that re-send long histories depend on cheap cache reads, and the comparison that matters is cached price per token on each vendor, not the headline input rate. The cross-vendor table on LLM API prices lines these up.

A pre-launch checklist

1. Set a spend limit below your tier cap on the Billing page.
2. Decide whether you need ZDR, and contact sales before you design around it.
3. Check the feature table for each API feature you use. Code execution, for instance, is outside ZDR.
4. Pin inference_geo if data location is a requirement, and test on a Claude 4.6 or later model.
5. Cache your system prompt and tool definitions, then measure the read rate.
6. Test your own prompts after any model switch, because the tokenizer changed.
7. Log the enforced_spend_limit_reached error code separately from ordinary 429s.

Our take

Anthropic's documentation is unusually explicit about retention. It names the models that need 30-day storage, says which features sit outside ZDR, and gives you a per-request region switch. That makes the vendor easier to assess than most. The weak points are the ones the docs also state: spend caps that stop service at $500 on the entry tier, and the top model being the one that cannot be run with zero retention by default. For most teams the sensible default is Sonnet 5.5, with Opus 5.5 reserved for the hard jobs and Haiku 5.5 for routing, and a cloud route if your security team already trusts one.

We have not run load or quality tests for this page. It reflects Anthropic's published documentation on 2026-10-31. Anthropic does not pay for placement on this page. If a provider buys a sponsored slot elsewhere, it is labelled and never changes ratings or order, as described in how we make money.

Related pages

Claude API pricing, LLM API prices, Claude models, Claude Code, Claude Agent SDK.