Claude API Pricing Explained: What Every Model Costs Per Million Tokens

Not affiliated with Anthropic. Claude Token Optimization is an independent site and audit service. Claude® and related product names are trademarks of Anthropic PBC. Plan prices, model rates, and usage limits change; always confirm on Anthropic’s own docs. [VERIFY: 2026-09-19]

Last updated: 2026-09-19. Rates below were checked against Anthropic’s published API pricing docs on that date. Model names and dollars change; re-check before you lock a budget. [VERIFY: 2026-09-19]

Searching for Claude API cost usually means you want one clean answer: what does each Claude model charge per million tokens, and how do cache and Batch change the bill? This page is that rate card.

If you need the foundation first, including what a token is, why input and output are priced differently, and why word count is a bad estimator, start with how tokens are counted. To estimate your cost before you send, open the calculator with your real monthly volume as soon as you have scanned the table below.

Current Claude API pricing per million tokens

Figures are USD per million tokens (MTok) on the first-party Claude API using standard global routing. [VERIFY: 2026-09-19]

[VERIFY: 2026-09-19]

Quick mental math: 1M input + 200k output on Sonnet 5 is about $2 + $2 = $4 before cache or Batch. Scale linearly. [VERIFY: 2026-09-19]

Sonnet 5’s $2 / $10 rates were announced as introductory through 2026-08-31 and are now documented as the standard price. The previously scheduled rise to $3 / $15 did not occur. [VERIFY: 2026-09-19]

How Claude API billing works

The Claude API has no flat API subscription for Messages traffic. You pay for tokens processed:

1. Input: system instructions, conversation history, tool schemas, file text in context, and the latest user message.
2. Output: the model’s generated reply, including structured tool-call payloads counted as generation under current metering. [VERIFY: 2026-09-19]
3. Prompt caching: optional cache writes at a premium and cache reads at a steep discount.
4. Batch API: optional asynchronous jobs at 50% off standard input and output. [VERIFY: 2026-09-19]

Seat plans such as Pro, Max, Team, and Enterprise are a different product surface. They meter chat and Claude Code capacity; they are not the same invoice as API token charges. If you are choosing between a seat and keys, see Pro or the API when that guide is live.

US-only inference through inference_geo: "us" on eligible Claude 4.6+ models applies a 1.1× multiplier across token categories, including cache. Global routing is the default and uses the table above. [VERIFY: 2026-09-19]

Current model rate breakdowns

Claude Haiku 4.5 pricing per million tokens

Haiku 4.5 is the cost floor most teams should price first for high-volume, latency-sensitive work. [VERIFY: 2026-09-19]

[VERIFY: 2026-09-19]

Haiku is where classification, routing, extraction, and good-enough drafting belong. If your default is Opus because it feels safer, you are often paying a 5× input premium for work Haiku finishes. For the decision rule, see when Haiku is enough.

Claude Sonnet 5 pricing per million tokens

Sonnet 5 is the production default for coding agents and most serious product workloads. [VERIFY: 2026-09-19]

[VERIFY: 2026-09-19]

At roughly 2× Haiku input and half of Opus input, Sonnet 5 is usually the right place to park an agent that must reason and call tools without Fable- or Opus-class spend. Unit price is not task price. A cheaper model that needs three retries can lose to a pricier model that finishes once, which is why cost per finished task matters more than a rate-card screenshot. [VERIFY: 2026-09-19]

Claude Opus 5 pricing per million tokens

Opus 5 is the heavyweight for complex agentic coding and hard enterprise reasoning. [VERIFY: 2026-09-19]

[VERIFY: 2026-09-19]

Opus makes sense when failure is expensive: multi-step refactors, ambiguous specifications, or reviews where a miss costs more than tokens. It is a poor default for FAQ bots, simple transforms, or anything you can evaluate automatically. Route hard cases up; do not leave the whole product on Opus just in case.

Fast mode, a research preview on Opus 5 and Opus 4.8, lists higher rates of about $10 / $50 input and output. It is a separate SKU decision, not the standard rate card above. [VERIFY: 2026-09-19]

Claude Fable 5.1 pricing per million tokens

Fable 5.1 sits at the top of the public rate card for long-running, high-stakes agent work. [VERIFY: 2026-09-19]

[VERIFY: 2026-09-19]

Note the cache-hit quirk: Fable 5.1 and Mythos 5.1 use a 0.025× cache-hit multiplier, or $0.25 / MTok, instead of the usual 0.1×. Do not copy Fable 5’s older $1 / MTok cache-hit figure into a 5.1 forecast. [VERIFY: 2026-09-19]

Mythos SKUs with limited availability list at Fable-like dollars. Omit them from default calculators unless you have access. [VERIFY: 2026-09-19]

Legacy Sonnet and Opus SKUs still on the API

Searchers still look up older generation names. As of 2026-09-19, these remain documented on the first-party API at the rates below. Retired Opus 4 / 4.1 and older Haiku SKUs are partner-only in many cases, so confirm before you budget them. [VERIFY: 2026-09-19]

Claude Opus 4.5, 4.6, 4.7, and 4.8: input $5 / MTok and output $25 / MTok, the same headline dollars as Opus 5 on the standard card. Cache rates are $6.25 for a 5-minute write, $10 for a 1-hour write, and $0.50 for a hit. Batch is $2.50 / $12.50. [VERIFY: 2026-09-19]

Claude Sonnet 4.5 and 4.6: input $3 / MTok and output $15 / MTok, higher than Sonnet 5’s now-standard $2 / $10. Cache rates are $3.75 for a 5-minute write, $6 for a 1-hour write, and $0.30 for a hit. Batch is $1.50 / $7.50. [VERIFY: 2026-09-19]

If your code still pins sonnet-4-5 or opus-4-6 IDs, you are not on the Sonnet 5 price. Migrating model IDs is often a larger cost lever than negotiating a discount. [VERIFY: 2026-09-19]

Prompt caching: write premium, read discount

Caching is how you stop re-paying full input price for the same system prompt, documentation, or tool schemas on every call.

[VERIFY: 2026-09-19]

Multipliers against base input for most models are 1.25× for a 5-minute write,2× for a 1-hour write, and 0.1× for a cache hit. Fable and Mythos 5.1 hits are 0.025×. A 5-minute write typically breaks even after about one successful read of the same block; a 1-hour write needs roughly two. [VERIFY: 2026-09-19]

When caching pays: large, stable prefixes reused across many requests. When it does not: one-off prompts, constantly mutating context, or write-heavy patterns with almost no hits. For the mechanics deep-dive, see prompt caching when published.

Batch API: 50% off when you can wait

[VERIFY: 2026-09-19]

Use Batch for overnight evaluations, backfills, report generation, and anything that can wait. Do not use it for interactive chat. Full limits, exclusions, and worked examples are on the Batch API pricing page. Batch discounts can stack with caching multipliers on eligible traffic; confirm stacking in Anthropic’s docs when you implement. [VERIFY: 2026-09-19]

What actually drives Claude API cost

Unit rates explain the price. Your workflow explains the bill.

1. Repeated context on every turn without caching.
2. The wrong model for the job, such as Opus or Fable on work Haiku or Sonnet can finish.
3. Verbose outputs and retry loops, including failed tools and try-again chains.
4. Synchronous paths for batchable work that could take the 50% Batch cut. [VERIFY: 2026-09-19]
5. US-only inference when global routing would have been fine, adding 1.1×. [VERIFY: 2026-09-19]
6. Legacy Sonnet 4.5 or 4.6 IDs at $3 / $15 when Sonnet 5 is $2 / $10. [VERIFY: 2026-09-19]

If the frustration is “we already pay more every month and quality is flat,” stop shopping rate cards alone. Structure beats another model upgrade.

Estimate your number, then decide what to cut

1. Open the cost calculator.
2. Enter monthly input and output, or prompts per day multiplied by average tokens.
3. Add a realistic cache-read share and Batch percentage if those apply.
4. Compare Haiku, Sonnet, and Opus on the same volume.
5. If the estimate is dominated by repeated prefixes or messy prompts, fix structure before you negotiate volume discounts.

Track subscription-style session limits separately on usage limits if you also live in Claude.ai. Those bars are not the same as a Messages invoice.

Get a written look at your own prompts

Rates explain the unit price. Your threads explain the multiplier.

If you have production prompts or a representative sample of chat and API turns and want a written map of where tokens go and what to change, get a written audit of your prompts. You can also read a sample report and learn who runs this.