AI Tools

Claude API pricing: what each model costs per million tokens

Published 2026-09-23 · Researched and price-checked on official sources

Claude API pricing now starts at $4 per million input tokens and $20 per million output tokens for Opus 5.5, which Anthropic cut from Opus 5's $5 and $25 when it launched the 5.5 family on September 22, 2026. Cache reads fell to $0.20 per million, and cached tokens are where agentic coding bills actually land.

Key facts

What matters

Claude Opus 5.5 bills $4 input and $20 output per million tokens, 20% below Opus 5, with cache reads at $0.20 per million instead of $0.50 (Anthropic pricing docs, September 23, 2026).
Sonnet 5 stays at $2 input and $10 output per million after Anthropic cancelled the $3/$15 increase it had scheduled for September 1, 2026.
The Batch API takes 50% off both directions, while Fast mode doubles Opus 5.5 to $8 input and $40 output and cannot be combined with Batch.
Claude 4.6 and later models bill the whole 1M-token context window at standard rates; OpenAI still publishes a separate, pricier long-context column.
DeepSeek's v4-pro lists $1.32 cache-miss input and $3.96 output per million at peak, so the cheapest credible alternative to Opus 5.5 is not from a US lab.

Claude API pricing by model, verified September 23, 2026

Claude API pricing is metered per million tokens (MTok) in both directions: what you send in, what the model writes back. Prompt caching, the Batch API, Fast mode and inference geography each apply their own multiplier on top of the base rates, which is why two teams with the same monthly token count can see bills that differ by 3x.

Claude Opus 5.5 landed on September 22, 2026 as the first model in the 5.5 family. Anthropic's launch post says input and output tokens are $4 and $20 per million, 20% less than Opus 5, cache reads are $0.20 per million, 60% less than Opus 5, and typical workloads cost about 40% less overall. The company also says Opus 5.5 writes output more than 30% faster than Opus 5.

Below the top of the family, Sonnet 5 anchors the value end at $2 input and $10 output per million. Anthropic's pricing page states the $2/$10 rate it announced as introductory pricing through August 31, 2026 is now the standard price, and the previously scheduled increase to $3/$15 on September 1, 2026 will not happen. Haiku 4.5 is $1 and $5. Fable 5.1, the flagship that sits above the Opus line, runs $10 and $50; Mythos 5.1 is listed at the same rate with limited availability.

Two housekeeping details matter for anyone pinning model IDs: Claude Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are retired except on Bedrock and Google Cloud, and the announcement says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, so the mid-tier rates are likely to move again soon.

The multipliers that decide your real Claude bill

Prompt caching. A 5-minute cache write costs 1.25x the base input rate and a 1-hour write costs 2x. Reads cost 0.1x base, dropping to 0.05x on Opus 5.5 and 0.025x on Fable 5.1 and Mythos 5.1. In Opus 5.5 numbers: $5 per million to write a 5-minute cache, $8 per million for a 1-hour cache, $0.20 per million to read it back. That spread is the point. Filling a 1-hour cache with 30 million tokens costs $240; reading those same 30 million tokens costs $6. Reuse pays, creation does not.

Batch API. Asynchronous work gets 50% off input and output, which puts Opus 5.5 at $2 and $10 per million and Sonnet 5 at $1 and $5.

Fast mode. A research-preview tier that speeds up output for Opus 5.5, Opus 5 and Opus 4.8 at double the standard rate: $8 input and $40 output per million. It covers the full context window, is available only on the first-party Claude API rather than AWS, Google Cloud or Azure, and cannot be used with the Batch API.

Data residency. Passing inference_geo with the US setting adds a 1.1x multiplier on every token category for Claude 4.6 and later models, cache reads included. Global routing is the default and stays at list price.

Long context. This one goes the other way. Claude 4.6 and later include the full 1M-token context window at standard pricing, and Anthropic's docs state that a 900k-token request is billed at the same per-token rate as a 9k-token request, with caching and batch discounts applying across the window.

What a month of agentic coding costs on each model

Take a realistic heavy-agent month: 30 million cached input tokens, 5 million fresh input tokens and 1 million output tokens. Priced at the published Claude rates, the same workload lands at:

Haiku 4.5: $13
Sonnet 5: $26
Opus 5.5: $46
Opus 5: $65
Fable 5.1: $107.50

The Opus 5.5 cut is worth about 29% against Opus 5 on this mix, close to Anthropic's own "40% less on typical workloads" claim once output volume is heavy. The bigger lever is routing. Send planning and refactors to Opus 5.5 and the mechanical edits to Sonnet 5, and the output line alone drops from $20 to $10 per million tokens of generated code. Anthropic publishes no router; the pricing docs simply give you the rates to build one.

Cache discipline compounds the same way. An agent that re-reads a 30-million-token context across a month of turns pays $6 for those reads on Opus 5.5, versus $15 on Opus 5 and $30 on gpt-6-astra at its cached-input rate. That line item is usually the largest one in a coding agent's bill, which is exactly why the Opus 5.5 cache-read cut from $0.50 to $0.20 per million matters more than the headline 20%.

Claude API pricing vs OpenAI and DeepSeek

OpenAI publishes Standard, Batch, Flex and Fast mode tiers on one page. Standard short-context rates are $10 input and $50 output per million for gpt-6-astra, $2 and $10 for gpt-6-sol, and $0.10 and $0.50 for gpt-6-luna. OpenAI still splits short and long context into separate columns, where gpt-6-astra rises to $20 input and $75 output; Anthropic charges its standard rate across the full window.

On the same 36-million-token workload used above, the metered totals are $5.94 to $11.88 on deepseek-v4-pro depending on peak or off-peak hours, $26 on gpt-6-sol, $46 on Opus 5.5, $107.50 on Fable 5.1 and $130 on gpt-6-astra. DeepSeek's own table lists cache hits at $0.044 per million on v4-pro at peak, against $0.20 for Opus 5.5 and $1.00 for gpt-6-astra.

The honest framing: Opus 5.5 is not competing on price per token against DeepSeek, which is roughly an order of magnitude cheaper on input. It is competing on what a token produces. If your workload is high-volume extraction, classification or summarisation where a smaller model is good enough, paying 4x to 8x more for frontier reasoning is money burned. If it is long-horizon agent work where a wrong edit costs a rollback and an hour, the token price is the smaller number.

Claude subscription or Claude API: which is cheaper for you

Anthropic's consumer pricing is Free at $0, Pro at $17 per month when billed annually ($200 up front) or $20 month to month, and Max from $100 per month with a choice of 5x or 20x Pro usage.

Against the $46 API month computed above, Pro at $17 or $20 looks like the better buy, and for interactive work it usually is. The Opus 5.5 launch shifted that line further: Anthropic used the release to raise five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and to give subscription users a rate-limit reset they can save and spend whenever they want. A developer who lives in Claude Code on one machine can often stay inside a Pro or Max plan and never touch metered tokens.

The split is about who is waiting. Interactive sessions with a human reading output fit a subscription, where the limit is a usage cap, not a bill. Overnight batch jobs, CI agents and anything that runs unattended belong on the API, where Batch pricing halves the rate and a usage cap cannot stop a pipeline at 2am.

The catches worth knowing before you budget

Fast mode is a premium, not a default. It doubles Opus rates, runs only on the first-party Claude API, and excludes the Batch discount, so the two cheapest levers and the speed lever cannot be pulled at once.

AWS and Azure buyers bill in CCUs. Anthropic's marketplace tables price the Claude Consumption Unit at a fixed $0.01, metered hourly, invoiced monthly, postpaid with no prepaid credits. Enterprise procurement that expected to pre-buy committed spend has to plan around that.

The flagship got relatively more expensive. Fable 5.1 stayed at $10 and $50 per million while Opus 5.5 dropped in under it, so the gap between the Opus line and the flagship is now 2.5x on output. If you were paying for the top model for every task, the 5.5 launch is the moment to re-check which calls actually need it.

Model retirement is a budget event. Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are retired on the first-party API, with Bedrock and Google Cloud as the remaining homes. Anything still pinned to those IDs needs a migration before the next price sheet.

At a glance

Claude model	Input	5m cache write	1h cache write	Cache hit	Output	Batch input / output
Opus 5.5	$4	$5	$8	$0.20	$20	$2 / $10
Opus 5, 4.8, 4.7, 4.6, 4.5	$5	$6.25	$10	$0.50	$25	$2.50 / $12.50
Sonnet 5	$2	$2.50	$4	$0.20	$10	$1 / $5
Haiku 4.5	$1	$1.25	$2	$0.10	$5	$0.50 / $2.50
Fable 5.1 (flagship)	$10	$12.50	$20	$0.25	$50	$5 / $25

Claude API list prices per million tokens, from Anthropic's pricing documentation, checked September 23, 2026.

FAQ

Is the Claude API free to use?

No. Claude's consumer app has a $0 Free plan, but API access is metered per token. The cheapest rate on Anthropic's current price list is Haiku 4.5 at $1 per million input and $5 per million output tokens, and Sonnet 5 at $2 and $10.

How much does Claude Opus 5.5 cost compared with Opus 5?

Opus 5.5 is $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5, a 20% cut. Cache reads fell harder, from $0.50 to $0.20 per million. The Batch API halves those to $2 and $10, while Fast mode doubles them to $8 and $40.

Does Claude API pricing go up for long context?

Not on Claude 4.6 and later models, which include the full 1M-token window at standard pricing; Anthropic's docs say a 900k-token request bills at the same per-token rate as a 9k-token one. OpenAI still publishes separate long-context columns, where gpt-6-astra costs $20 input and $75 output per million.