Claude API Pricing in 2026

The full Claude API rate card, the modifiers that change it, and where teams get the math wrong.

Claude API pricing looks like the simplest rate card in the industry. Four current models, one number for input, one for output, and a batch discount that halves both.

Then you try to forecast with it, and it comes apart.

In Claude API pricing the list price is only one of about six things deciding what you pay. A newer tokenizer counts the same document differently. Two of Anthropic’s own distribution channels bill you in a currency that is not dollars. And the largest cost lever on the platform, the 1M token context window, is free.

Every figure here was verified on September 21, 2026 against Anthropic’s pricing documentation. For a number against your own volumes, our free LLM cost calculator covers 100+ models across 10 providers, and our Anthropic pricing reference breaks the lineup down by model type.

Claude API pricing: the full rate card

All figures are USD per million tokens.

Model	Input	Cache write (5m)	Cache write (1h)	Cache read	Output
Claude Fable 5.1	$10	$12.50	$20	$0.25	$50
Claude Opus 5	$5	$6.25	$10	$0.50	$25
Claude Sonnet 5	$2	$2.50	$4	$0.20	$10
Claude Sonnet 4.6	$3	$3.75	$6	$0.30	$15
Claude Haiku 4.5	$1	$1.25	$2	$0.10	$5
Claude Opus 4.1 (Retired)	$15	$18.75	$30	$1.50	$75

Claude Mythos 5.1 matches Fable 5.1’s rates but is limited availability, gated behind Anthropic’s approval-based Project Glasswing, so it is not a line you can plan around.

Three things in that table decide most budgets.

Sonnet 5’s price is settled. Anthropic launched $2 and $10 as introductory pricing due to expire August 31, 2026, with an increase to $3 and $15 from September 1. That increase was cancelled and $2 and $10 is now the standard price. A lot of published guidance still has this wrong.

The legacy Opus rate is triple. At $15 and $75, Opus 4.1 costs three times the current flagship. Opus 4.1 is still reachable on both Amazon Bedrock and Google Cloud, while Opus 4 is retired everywhere except Google Cloud. Opus 4.5 through 4.8 all sit at the current $5 and $25 and are fine to stay on.

Newer does not mean cheaper. Sonnet 5 undercuts Sonnet 4.6, but Opus 5 matches Opus 4.8 exactly, Fable 5.1 matches Fable 5, and Haiku 4.5 at $1 and $5 is dearer than the retired Haiku 3.5 at $0.80 and $4. Only one tier came down.

The tokenizer catch that undoes a price cut

Anthropic’s Sonnet 5 release notes state that Claude 4.7 and later use a newer tokenizer producing approximately 30% more tokens for the same text, so the cost of an equivalent request does not drop in proportion to the per-token price.

The multiplier is not flat. Anthropic’s launch post is reported as putting the range at roughly 1.0x to 1.35x by content type, and independent testing by Simon Willison measured token counts, not prices, at 1.42x on English, 1.33x on Spanish, 1.27x on Python, and essentially unchanged on Simplified Mandarin. Applying those multipliers to the published rates:

Content type	Token multiplier	Effective Sonnet 5 input	Real saving vs Sonnet 4.6 at $3
Simplified Mandarin	~1.0x	$2.00	33%
Python code	1.27x	$2.54	15%
Spanish	1.33x	$2.66	11%
English prose	1.42x	$2.84	5%

So a move that reads as a 33% price cut delivers roughly 5% on an English-heavy workload.

The harder case is where the list price does not move at all. Opus 4.6 and Opus 5 are both $5 and $25, but Opus 4.6 uses the old tokenizer and Opus 5 the new one, so on English text the same document effectively costs around $7.10 per million. Identical rate card, roughly 40% more per page.

This is not a hidden fee, and a model that finishes a task in fewer turns can still be cheaper overall. The point is narrower: you cannot compare Claude models on list price across a tokenizer boundary. Run your own corpus through the token counting endpoint before migrating.

Claude API pricing and the cache TTL most teams pick wrong

Cache reads cost 0.1x the base input rate. Writes cost 1.25x for the five-minute cache or 2x for the one-hour cache, and Anthropic’s guidance is that five minutes pays for itself after one read, one hour after two.

The detail that decides it sits in the column heading: that 0.1x rate covers cache hits and refreshes. A read restarts the clock at read price.

So TTL choice is not about traffic volume, it is about the longest gap between requests sharing a prefix. Under continuous traffic the five-minute cache is refreshed by its own reads and never needs rewriting. The one-hour cache earns its 2x when quiet periods regularly run past five minutes but rarely past an hour. Count your idle gaps before choosing.

Fable 5.1 and Mythos 5.1 are the outlier, pricing cache reads at 0.025x rather than 0.1x, so heavy-cache workloads narrow the gap to cheaper models considerably.

What Claude API pricing does not charge extra for

The full 1M token context window is billed at standard rates. A 900k-token request bills at the same per-token rate as a 9k-token request on Claude 4.6 and later, with caching and batch discounts applying across the whole window. There is no long-context tier and no threshold multiplier, so if you have been architecting around a context cliff, you are solving a problem Claude does not have.

Web fetch is free, beyond token costs for what it pulls in. Code execution is free when paired with web search or web fetch. Used alone it bills by execution time with a five-minute minimum, after 1,550 free hours a month per organisation, then $0.05 per hour per container. One trap: if files are included in the request, execution time bills even if the tool is never called.

The Claude API pricing modifiers that stack

Modifier	Effect	Notes
Batch API	0.5x input and output	Asynchronous. Not available with fast mode
Fast mode	$10 / $50 per MTok	Research preview, Opus 5 and Opus 4.8 only
US-only inference	1.1x all token categories	inference_geo: “us”, Claude 4.6 and late
Regional endpoints on Bedrock or Vertex	10% premium	Over global endpoints

Data residency is the quiet one, because 1.1x applies to everything, so cache reads and writes get the premium too.

Tools add cost before your prompt does anything. A tool-enabled request carries 286 system prompt tokens on Opus 5 but 496 on Haiku 4.5, so the cheaper model has the heavier overhead, and individual toolsets stack on top: roughly 4,500 tokens for computer use and 6,600 for browser use. Web search costs $10 per 1,000 searches plus tokens, with failed searches unbilled. Managed Agents add $0.08 per session-hour on top of tokens, billed only while a session is running, so idle time is free.

Buying Claude through AWS or Azure changes your bill format

Claude API pricing is identical here, but the invoice is not. Claude Platform on AWS and Claude in Microsoft Foundry bill through their marketplaces using Claude Consumption Units. Anthropic rates usage in USD at standard rates, applies any discount, then converts to CCUs at a fixed $0.01 each.

The rate you pay is the same. What changes is what you can see.

Anthropic’s docs are explicit that your AWS bill shows a single CCU line item, and that AWS Cost Explorer and Azure Cost Management show aggregated CCUs. The breakdown by model and feature lives in the Claude Console, not your cloud cost tool. CCU billing is also postpaid only, and negotiated discounts appear as fewer CCUs metered rather than a lower price, so the discount is invisible in the line item. Our guide to Anthropic cost monitoring covers the tracking options.

Claude API pricing in practice: a worked example

A document pipeline handling 200,000 requests a month. Each sends 5,000 input tokens, 4,000 of them a stable cached prefix, and returns 800 output tokens.

Component	Fable 5.1	Opus 5	Sonnet 5	Haiku 4.5
Uncached input	$2,000	$1,000	$400	$200
Cached input	$200	$400	$160	$80
Output	$8,000	$4,000	$1,600	$800
Monthly total	$10,200	$5,400	$2,160	$1,080

This assumes a warm cache kept alive by refreshes, so cache-write cost is treated as negligible. Traffic with gaps longer than your TTL will add cold-start writes.

Note the cached input row: Fable 5.1 costs half what Opus 5 does on cached reads despite being twice the price everywhere else. Three more numbers from the Sonnet 5 column. Without caching it costs $3,600, so the cache saves 40%. On Batch it halves to $1,080, exactly what Haiku 4.5 costs at standard rates. And switching to Opus 5 multiplies the bill by 2.5 for identical traffic. You can model your own volumes in the LLM cost calculator.

Claude pricing vs Claude API pricing: two different bills

Everything above is Claude API pricing. The subscriptions at claude.com/pricing, covering Free, Pro, Max, Team, and Enterprise, are billed separately per seat.

Three nuances catch teams out. A subscription does not come with API credits and API spend does not draw down a plan. Usage beyond plan limits is purchased as credits billed at standard API rates, so the underlying meter is the same one above. And Enterprise is a seat cost plus API usage rather than an all-in number, which makes the quoted per-seat figure a floor rather than a forecast.

New API accounts are also widely reported to receive a small one-time credit, commonly cited as around $5. Anthropic does not publish the amount, so treat it as anecdotal.

How to reduce Claude API costs

• Move off retired models. Opus 4.1 and Opus 4 at $15 and $75 cost three times Opus 5. This is a model string change.
• Re-tier by measured quality. The spread from Fable 5.1 to Haiku 4.5 is 9.4x on the workload above.
• Send anything asynchronous through Batch. Half price, and it stacks with caching.
• Match cache TTL to your idle gaps, not your traffic volume.
• Audit tool overhead. Toolset definitions add thousands of tokens per request.
• Measure tokens on the model you deploy, since the tokenizer boundary makes list-price comparisons unreliable.

How Economize tracks Claude spend

Those levers are engineering work. Knowing which team, feature, or customer triggered the spend is not, and it is the part Anthropic’s console cannot finish once CCU billing puts a single aggregated line on your cloud invoice.

Economize brings usage-based Claude and OpenAI costs into the same dashboards as your AWS, Azure, and GCP spend.

• AI and cloud spend in one view. Anthropic and OpenAI costs sit alongside AWS, Azure, and GCP in the same reports, so marketplace-billed Claude spend does not vanish into a line nobody can break down.
• Allocation without a re-tagging project. Virtual tagging applies rules across existing tags, including shared and untaggable costs, turning an API key into a customer, feature, or team.
• Anomaly detection with root cause analysis. Catches what a flat invoice hides, a rate change absorbed by a usage increase.
• An AI FinOps agent. Anyone can query cost data directly and act on it instead of filing a request.
• Published, flat pricing. Free up to $100,000 a month in combined cloud and AI spend, $249 up to $250,000, and enterprise from $2,499 with SSO, RBAC, and audit logs.

Setup is agentless and takes about five minutes. Economize is SOC 2 certified, and customers including Hasura, DeepSource, Capchase, and Apna report savings up to 30%.

Frequently Asked Questions

Claude API pricing will move again. Try the free LLM cost calculator to size your own spend, or connect your accounts to Economize and see AI and cloud costs in one place.