Claude Pricing Explained (August 2026): Subscription vs API Costs, and How to Reduce Them
- il y a 2 jours
- 5 min de lecture
Dernière mise à jour : il y a 11 minutes
Prices below were verified on Anthropic's official pricing pages on August 10, 2026. Anthropic changes them regularly; check the sources at the end before you commit budget.
Anthropic sells access to Claude in 2 ways. A subscription is billed per person per month, like Microsoft 365. The API is billed per token consumed, like an AWS account. The 2 models answer different questions: a subscription prices a person, the API prices a workload. Companies running Claude at scale usually end up paying for both, often without having decided to. All prices below are Anthropic list prices as of August 2026.
Subscription pricing: a flat fee per person
A subscription gives a named person access to Claude through the chat interface, the mobile and desktop apps, Claude Code in the terminal and Cowork.
Plan | Price | What you get |
|---|---|---|
Free | $0 | Chat on web, mobile and desktop, with modest usage limits |
Pro | $17/month billed annually, $20 billed monthly | Higher limits, Claude Code, Claude Cowork, all models |
Max 5x | $100/month | 5x the Pro usage allowance |
Max 20x | $200/month | 20x the Pro usage allowance, priority access at peak times |
Team Standard | $20/seat/month billed annually ($25 monthly) | Central billing, admin controls, more usage than Pro |
Team Premium | $100/seat/month billed annually ($125 monthly) | 5x the usage of a standard seat |
Enterprise | $20/seat billed annually, plus usage at API rates | SSO, SCIM, audit logs, compliance API |
The word "limits" matters. The fee is flat but the usage is not unlimited: every plan carries a usage ceiling that resets on a rolling window, and heavy users reach it. And the Enterprise plan is no longer a pure seat price. Anthropic lists it at $20 per seat plus usage billed at API rates, so an Enterprise contract already behaves partly like an API account.
Claude API pricing: pay per token
The API is for software, not people. Your application sends a request, Claude answers, and Anthropic meters both directions in tokens. A token is roughly 3/4 of an English word, for example: "cloud cost optimization" is 4 tokens. You pay one rate for input tokens (what you send, including documents and conversation history) and a higher rate for output tokens (what Claude writes back).
Current base rates per million tokens (MTok):
Model | Input | Output |
|---|---|---|
Claude Fable 5 | $10 | $50 |
Claude Opus 5 | $5 | $25 |
Claude Sonnet 5 | $2 | $10 |
Claude Haiku 4.5 | $1 | $5 |
Model choice sets your unit economics. The spread between Haiku and Fable is 10x on both input and output, and a forecast built on Sonnet's $2/MTok input gains 50% on that line item in 3 weeks.
Base rates are only the anchor. The Batch API, for jobs that can wait up to 24 hours, cuts both prices by 50%. Prompt caching bills repeated context at 0.1x the input price, after a cache write charged at 1.25x or 2x depending on cache duration. US-only inference adds a 1.1x multiplier. Web search costs $10 per 1,000 searches on top of token costs. And Fast mode, a research preview on Opus, doubles the rate for lower latency. These modifiers stack, which is why 2 teams calling the same model can pay very different effective prices.
Claude models from version 4.7 onward use a new tokenizer that produces roughly 30% more tokens for the same text than older Claude models did. A tokenizer is the component that splits text into the units you are billed for, and vendors revise it for engineering reasons, not pricing ones. The practical effect is the same either way: the same document costs more tokens to process on a 4.7+ model than it did on Sonnet 4.5, before any rate change. Per-token price comparisons across model generations are unreliable unless you compare cost per document, not cost per token.
Subscription vs API: which one fits your use case
Buy subscriptions when humans use Claude interactively: an analyst drafting reports, a developer in Claude Code, a consultant preparing client material. The cost is capped by design. When a Pro user hits the allowance, they wait for it to renew; the seat stays at $20 unless someone deliberately buys extra usage credits. Our rule from client work: an individual developer needs Max only after hitting Pro limits repeatedly in a normal working week.
Predictability is the second argument. Finance can budget 40 seats at $100. Budgeting 40 developers' token consumption is a forecasting exercise with an error bar.
Use the API when software calls Claude. A support bot answering 50,000 tickets a month, a document pipeline classifying contracts, an agent monitoring your cloud spend overnight. A subscription cannot serve these workloads, because it authenticates a person rather than an application.
Here volume is machine-driven, so cost scales with traffic rather than headcount, and there is no cap unless you configure one. This is where FinOps discipline applies from day one: A chatbot that doubles its traffic doubles its bill without anyone approving anything. Better tag API keys by team and workload, set spend alerts and forecast in tokens rather than seats.
The boundary case is the developer team. Claude Code works with both a subscription and an API key. On a subscription the cost is fixed per developer; on an API key it is variable and, for a developer using it daily, higher. Default to subscriptions for humans and API keys for CI pipelines and automation.
How to reduce Claude costs: 5 ways to OptimNow the bill
Match the model to the task. The spread between Haiku 4.5 ($1/$5) and Fable 5 ($10/$50) is 10x on both input and output. Classification, extraction and routing rarely need the top model.
Move non-urgent work to the Batch API. Anything that tolerates a delay of up to 24 hours (nightly classification, report generation, evaluation runs) costs 50% less. This is the simplest discount in the catalog and the most often ignored.
Cache your prompts. If your requests share a long common prefix (a system prompt, a policy document, a knowledge base extract), prompt caching bills cached input at 10% of the base rate. For RAG applications this is routinely the largest single saving available.
Put humans on subscriptions, machines on the API. A human on a metered API key is an uncapped cost center, and a batch job running behind a Max seat hits usage limits that were sized for a person. Sort the population before optimizing anything else.
Turn on the governance features before you need them. Since July 2026, Claude Enterprise admins get a spend dashboard, per-role model entitlements and threshold alerts at 75% and 90% of a spend limit. On the API side, set workspace spend limits and review usage by key monthly. Configure the alerts before the first overrun, not after it appears on an invoice.
Where to start
Make an inventory before touching any setting: who holds a paid seat, which workloads hold an API key and which model each one calls. The list fits on one page for most organizations below a few hundred users, and it answers the question an invoice cannot: who spends what, through which billing model. If the inventory shows a workload calling Fable 5 for a task Haiku 4.5 could handle, you have found your first saving, and it is a one-line change.
The FinOps skill OptimNow uses to track Anthropic billing is open source under CC BY-SA: github.com/OptimNow/cloud-finops-skills. Download it, load it into your own Claude and run the same analysis on your invoice.
.png)