Claude API Pricing 2026: What Tokens Actually Cost

Claude API costs from 1 USD per million input tokens and 5 USD per million output tokens (Haiku 4.5) up to 10 and 50 USD on Fable 5. Sonnet 5 runs at a promotional 2 / 10 USD until 31 August 2026 and moves to its standard 3 / 15 USD from 1 September 2026. Figures as of August 2026.

Billing is consumption-based: you pay for the tokens the model read and the tokens it wrote. A token is a fragment smaller than a word, so one page of a document sits in the range of a few hundred tokens. No subscription, no minimum, and no ceiling either unless you set one.

Current Claude API price list (August 2026)

Rates per million tokens, standard tier:

The Sonnet 5 promotion is dated: 2 / 10 USD until 31 August 2026, then 3 / 15 USD from 1 September 2026. If you are budgeting a quarter, use the September rate so your plan does not rest on a price that expires inside it.

Three things worth reading out of that table. The gap between Haiku and Fable is tenfold, so one model for everything is usually the most expensive option available. Output costs five times input on every model, so answer length moves the bill more than prompt length. Three models carry a one-million-token window while Haiku carries 200,000, which decides the matter whenever you feed in long case files.

Model prices change faster than service price lists, so verify the rates at the source in the Anthropic price list before any serious calculation.

How do you calculate a monthly Claude API bill?

You need three numbers: case volume, average tokens per case, and the model rate. Input and output are counted separately.

Monthly cost (USD) =
  cases per month
  x (input tokens per case / 1,000,000) x input rate

+ cases per month
  x (output tokens per case / 1,000,000) x output rate

Do not guess the tokens per case. Take a few dozen real cases, run them through the model once, and read the usage figures off the API response. That sample is the only number worth multiplying by volume.

Then add headroom. The system prompt, the conversation history, and any attached documents enter the input on every single call, and an agent that takes several steps per case pays for input once per call. This is where first estimates usually turn out too low.

Why is output more expensive than input?

Because the model generates it token by token, while input is processed in one pass. The practical consequence is that the cheapest processes read a lot and write a little: triaging a ticket, pulling fields off an invoice, deciding a route. The expensive ones produce long text: reports, lengthy customer replies, whole-document translation.

That gives you a design rule. If the answer goes to a person, ask for structure and a length cap. If the output feeds another system, force a compact format. Halving the average response does more for the bill than switching to a cheaper model.

How do you cut the Claude API bill?

Four levers, largest first:

• Model choice per step. Haiku 4.5 for classification and routing, Sonnet 5 for routine work, Opus 5 or Fable 5 only for cases that genuinely need them. A process where 80 percent of traffic runs on the 1 USD model and the expensive one handles exceptions costs a fraction of the same process on a single flagship.
• Prompt caching. A cache hit costs roughly 10 percent of the input rate. If every case carries the same pack of instructions, rules, and examples, this is the fastest saving on the bill.
• Batch API. Batch processing halves the rates. It fits anywhere the answer does not need to arrive within a second: overnight document processing, database enrichment, reporting.
• Daily and monthly limits. Set them before the system reaches production, and decide in advance what happens when the budget is hit: queue the work, fall back to a cheaper model, or alert a human.

Optimised bill ≈
  (input cost x (1 - cache share) + input cost x cache share x 0.1)
+ output cost
  all of it x 0.5 when the traffic runs through the Batch API

The first two levers are process design work, and no console switch delivers them. That is why two builds of the same process can differ several times over on the monthly invoice.

Claude through the clouds: Bedrock, Vertex, Foundry

Claude is also available through AWS Bedrock, Google Vertex AI, and Microsoft Foundry (generally available since 29 June 2026). Each route has its own price list and its own contractual terms, so do not carry Anthropic rates across to them automatically. The route also decides data residency: guaranteed EU processing comes today from Bedrock with the EU inference profile, while the direct Anthropic API processes requests in the US or globally. We unpack that in our analysis of Claude and GDPR.

Tokens are not the whole cost

In a typical build the API bill is the smallest line. The rest is work: process analysis, integrations, boundaries and escalation, testing, launch, and maintenance. Our lines look like this:

Architecture adds its own lines. Claude Managed Agents (public beta since 8 April 2026) bills managed agent infrastructure at 0.08 USD per hour plus tokens. Hosting, monitoring, and post-launch changes are priced individually. The full price list is on the pricing page, with wider cost context in the guide to AI agent implementation costs and in deploying Claude in your company.

When the API is the wrong purchase

Not every case deserves its own integration.

• Small, irregular volume. If this is a handful of people working with documents and an inbox, a subscription is cheaper and faster to start. The plan comparison sits in our piece on Claude Team and Enterprise.
• A process without rules. If you cannot describe what should happen in a typical case and in an exception, the API will not supply that knowledge. Sort the process out first.
• No business owner. A metered system needs someone who reads the invoice and judges answer quality. Without that person the cost grows quietly.
• A one-off task. For an analysis you run once, the Claude app is enough. Integration earns its keep on repetition.

To be plain about it: if the scan shows that a team plan and two hours of training will do, we say so.

How to start

1. Pick one process and count its monthly volume.
2. Measure tokens across a few dozen real cases and put them into the formula above.
3. Set the result against the labour the process consumes today, and decide whether it is worth building.

Book a free process scan and you will get a written takeaway with a model-cost estimate and the scope of the build. See also the AI agents service and the pricing page.

Frequently asked questions

How much does a Claude token cost? Rates are quoted per million tokens, priced separately for input and output. As of August 2026 that is 1 / 5 USD for Haiku 4.5, 2 / 10 USD for Sonnet 5 under its promotion until 31 August 2026 (3 / 15 USD afterwards), 5 / 25 USD for Opus 5, and 10 / 50 USD for Fable 5. Prices move quickly, so verify them in the Anthropic price list.

Which Claude model is cheapest? Haiku 4.5, at 1 USD per million input tokens and 5 USD per million output tokens, with a 200,000-token context window. It suits classification, routing, and short extractions. For long documents or hard reasoning, pick Sonnet 5, Opus 5, or Fable 5, each with a one-million-token window.

What is prompt caching in the Claude API? It stores the repeated part of your prompt (instructions, documentation, examples) on the provider side. A cache hit costs roughly 10 percent of the normal input rate. For an agent that attaches the same rule pack to every case, this is usually the single largest saving available on the bill.

How do I estimate Claude API cost before starting? Take one case type, measure input and output tokens across a sample of a few dozen real cases, then multiply by monthly volume and the model rate. Count input and output separately, because output costs several times more. Add build and maintenance costs, since tokens are usually the smallest line in the budget.

Related articles

• Deploying Claude in your company: 2026 costs and scope
• Claude Team vs Enterprise: pricing and features 2026
• Is Claude GDPR compliant in 2026?
• AI agent implementation cost: 2026 pricing guide
• Integrating ChatGPT and Claude with business systems