Anthropic’s pricing for Claude spans both end-user plans and usage-based API billing, but buyers evaluating the developer platform need to focus on model tier, token category, caching, and deployment path. For B2B teams, the main budget question is not just which model to use, but how prompt structure, endpoint choice, and billing route change total spend.
What it is and who it’s for
Claude API pricing is Anthropic’s usage-based pricing model for developers and enterprise teams building on Claude models. The platform docs state prices are in USD and organize charges by base input tokens, cache writes, cache hits and refreshes, and output tokens across active model families including Claude Fable 5, Claude Mythos 5, Claude Opus, Claude Sonnet, and Claude Haiku.
This guide is most relevant for teams buying AI capacity for product features, internal tools, or agent workflows rather than for individual chat use. Anthropic’s public plans page separates consumer and workspace plans from API-style enterprise billing: Free, Pro, Max, Team, and Enterprise exist for Claude users, while the Enterprise plan for large businesses combines a per-seat price ($20/seat) with usage billed at API rates that scales with model and task.
How it works
Claude API pricing starts with tokens, which are the units Anthropic uses to meter model usage. The pricing docs break token charges into several categories rather than a single rate: base input tokens, 5-minute cache writes, 1-hour cache writes, cache hits and refreshes, and output tokens.
Prompt caching is one of the main mechanics buyers need to understand. Anthropic says prompt caching reduces costs and latency by reusing previously processed prompt prefixes. A prompt prefix is the reusable earlier portion of a request, and Anthropic says caching can cover the full prefix across tools, system, and messages up to the cache_control block.
There are two enablement modes. Automatic caching uses a top-level cache_control field and is recommended for multi-turn conversations, while explicit cache breakpoints place cache_control on individual content blocks for finer control. By default, cache lifetime is 5 minutes, and Anthropic also offers a 1-hour cache duration at additional cost.
The pricing logic for caching is multiplier-based. Anthropic lists 5-minute cache writes at 1.25x the base input price, 1-hour cache writes at 2x, and cache hits at 0.1x. Cache hits are refreshed for no additional cost when reused.
Pricing and cost considerations
For buyers comparing model families, the first cost lever is model tier. Anthropic’s docs place Claude Fable 5 and Claude Mythos 5 at the top pricing tier, while Claude Haiku 4.5 is the lowest active model price listed.
The second lever is token type. Anthropic does not use one all-in rate per model; input, output, and cache actions are priced separately. A team that repeatedly sends large shared instructions or tool definitions may pay meaningfully different amounts depending on whether those prefixes are cached and then reused as hits.
The third lever is timing and model lifecycle. Claude Sonnet 5 launches at introductory pricing that steps up to a higher standard rate on a set date, shown in the schedule below. Anthropic also notes that Claude Opus 4.1 is deprecated, Claude Opus 4 is retired except on Google Cloud, Claude Sonnet 4 is retired except on Bedrock and Google Cloud, and Claude Haiku 3.5 is retired except on Bedrock and Google Cloud.
For standard token pricing, the currently documented rates break down as follows:
| Model family | Base input | 5-minute cache write | 1-hour cache write | Cache hit / refresh | Output |
|---|---|---|---|---|---|
| Claude Fable 5 / Claude Mythos 5 | $10 / MTok | $12.50 / MTok | $20 / MTok | $1 / MTok | $50 / MTok |
| Claude Opus 4.8 / 4.7 / 4.6 / 4.5 | $5 / MTok | $6.25 / MTok | $10 / MTok | $0.50 / MTok | $25 / MTok |
| Claude Sonnet 4.6 / 4.5 / 4 | $3 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok | $15 / MTok |
| Claude Haiku 4.5 | $1 / MTok | $1.25 / MTok | $2 / MTok | $0.10 / MTok | $5 / MTok |
Anthropic also documents a separate Sonnet 5 schedule:
| Model | Base input | 5-minute cache write | 1-hour cache write | Cache hit / refresh | Output |
|---|---|---|---|---|---|
| Claude Sonnet 5 introductory pricing through August 31, 2026 | $2 / MTok | $2.50 / MTok | $4 / MTok | $0.20 / MTok | $10 / MTok |
| Claude Sonnet 5 standard pricing starting September 1, 2026 | $3 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok | $15 / MTok |
Two additional cost adjustments can change the effective rate. Anthropic says regional and multi-region endpoints for Claude 4.5 models and beyond carry a 10% premium over global endpoints, and US-only inference for Opus 4.6, Sonnet 4.6, and later models uses a 1.1x pricing multiplier.
Anthropic also says prompt caching multipliers stack with Batch API discount and data residency.
How to choose
The cleanest way to choose among Claude API pricing tiers is to start with workload shape rather than headline model names. If the application reuses long system prompts, shared instructions, or tool schemas across many requests, prompt caching should be part of the cost model from the start because Anthropic prices cache writes and cache hits separately from ordinary input.
Teams that need the lowest listed active model price may start with Claude Haiku 4.5. Teams that need access to the highest-priced tiers documented by Anthropic will be looking at Claude Fable 5 or Claude Mythos 5. Teams weighing Anthropic against the other major API provider can compare these tiers with our OpenAI API pricing guide.
Operational constraints should be part of the selection process as well. If the deployment must use regional, multi-region, or US-only inference settings, buyers should include those endpoint multipliers in procurement review rather than treating them as minor implementation details.
A separate decision is whether the buyer needs user-facing Claude workspace features instead of, or alongside, API usage. Anthropic’s plans page shows that Team includes central billing and administration, SSO, admin controls for connectors, enterprise deployment for desktop, and no model training by default.
Limitations and gotchas
Claude API pricing has several edge cases that can trip up budget planning if teams only look at base token rates.
First, partner billing can use a different unit. On Claude Platform on AWS and Claude in Microsoft Foundry, Anthropic says usage is billed through Claude Consumption Units rather than directly through the token-rate table, at $0.01 per CCU with hourly metering, monthly invoices, and postpaid arrears only.
Second, API economics can improve with caching, but caching changes the request design. Anthropic says automatic caching moves the cache point forward as conversations grow, while explicit cache breakpoints give fine-grained placement control through cache_control on individual content blocks.
Third, enterprise buyers should separate workspace pricing from API economics. Anthropic’s Enterprise plan on the main pricing page combines a $20 per-seat price with usage billed at API rates that scales with model and task, and adds spend limits, role-based access, SCIM, audit logs, compliance API, custom retention controls, network-level access control, IP allowlisting, and a HIPAA-ready offering.
FAQ
Can prompt caching be used with Zero Data Retention?
Yes. Anthropic states prompt caching is eligible for Zero Data Retention when the organization has a ZDR arrangement.
How large is the Batch API discount? Anthropic prices Batch API usage at a 50% discount on both input and output tokens compared with standard rates.
Which payment methods does Anthropic list for paid Claude plans?
Anthropic lists credit card, ACH, and invoicing/net terms.
Sources
- Anthropic, Plans & Pricing | Claude by Anthropic
- Anthropic, Pricing - Claude Platform Docs
- Anthropic, Prompt caching - Claude Platform Docs
