Anthropic API Pricing (2026): Every Claude Model Per Million Tokens, and What a Real Workload Costs

TL;DR: The Rate Table Is the Easy Part

Every "Anthropic API pricing" page on the internet is a copy of Anthropic's table with ads around it. The table is below, read straight from Anthropic's pricing page on the date at the top of this post, because I'd rather you get the number from me than from a page that hasn't been updated since spring. But the table isn't the bill. The bill is the table times your token mix, times whether you cache, times which generation of model you're on — and those three multipliers move the number more than the choice between vendors does. I run production workloads on this API for clients; here's the whole picture.

Anthropic API pricing per million tokens (September 2026)

Opus 4.1 and Opus 4 are still listed at $15 / $75, but Anthropic has retired both on its own API (Opus 4.1 survives on Bedrock and Google Cloud, Opus 4 on Google Cloud only); if you're still on either there, you're paying almost four times the Opus 5.5 price for an older model. The Batch API is 50% off input and output on every model (Opus 5.5 batches at $2 / $10). Fast mode, a research preview on Anthropic's own API only, is a premium rate: $8 / $40 on Opus 5.5 and $10 / $50 on Opus 5 and Opus 4.8. Note the two cache exceptions in the table: Opus 5.5 cache hits are priced at 5% of input and Fable 5.1's at 2.5%, where every other model uses 10%. Prices are Anthropic's published US list rates as of September 2026; the cloud marketplaces (AWS, Microsoft Foundry) bill in their own units. Check the source page before you sign anything with a zero on the end.

September 2026 update: Opus 5.5, and what's retiring

Anthropic launched Claude Opus 5.5 on September 22, 2026 at $4 / $20 per million tokens, 20% under Opus 5's $5 / $25, with cache hits at $0.20 instead of $0.50. Anthropic says it performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5 once you count the fewer tokens it uses per task. Its model guide now says to start with Opus 5.5 for most workloads, and lists Opus 5 as a legacy model. The launch details, plan access and the Fable comparison are on Claude Opus 5.5.

Two dates to put in your calendar. Anthropic lists Haiku 4.5's retirement as not sooner than October 15, 2026; it's still active and no retirement is scheduled yet, but that's a floor, not a promise. And the rest of the 5.5 family is arriving: Sonnet 5.5 shipped September 28 at exactly Sonnet 5's rates in every column ($2 / $10, cache hits $0.20, batch $1 / $5), so Anthropic's "up to 30% less to run" for it comes entirely from using fewer tokens, not from the rate card. If a model ID is hard-coded anywhere in your stack, now is a good week to make it a setting.

October 2026 update: Haiku 5.5 shipped October 7 at $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens, 90% under Haiku 4.5, and $0.50 / $2.50 for prompts over that line. Anthropic puts the average saving at around 75% once longer prompts and its new tokenizer are counted. The same day it halved Sonnet 5.5's cache hits to $0.10, which it says makes Sonnet 5.5 about 20% cheaper on most agentic work. Two worked workloads, and why the 100K line should change how you chunk long documents, are on Claude Haiku 5.5. And if you pay for Max or Team, part of this bill may now be covered: those plans include a monthly API credit of $100 to $500, explained in Claude API credits.

Is the Claude API free?

Not as a tier. Anthropic's pricing page says new users get a small amount of free credits to test the API; after that it's pay-per-token. The thing that's free is the consumer chat at claude.ai, which is a different product with different limits. The practical version of "free": at Haiku 4.5's $1 per million input tokens, a thousand test requests with a 2,000-token prompt cost about $2. Evaluation is not where the money goes.

The price cut that was supposed to expire, and didn't

Sonnet 5 launched at $2 / $10 as "introductory pricing through August 31, 2026," with a scheduled increase to $3 / $15 on September 1. As of September 2026, Anthropic's page says the increase will not occur — $2 / $10 is now the standard price. That's a third off the previous Sonnet generation's rate for the newer model, and it changes the default I give clients: Sonnet 5 is now the cheapest mid-tier frontier model I'm aware of, and for most business workloads it's the one to start on. If someone quotes you Sonnet at $3 / $15, they're reading the old row.

The tokenizer gotcha: 4.7 and later count ~30% more tokens

Here's the line on Anthropic's page that nobody puts in their pricing roundup: Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text. The exact figure depends on the content. Sonnet 4.6 and earlier use the previous tokenizer. Why it matters: a per-token price comparison across generations understates the newer models' cost by roughly that much. Sonnet 5 at $2 with 30% more tokens bills like Sonnet 4.6 at about $2.60 — still cheaper than 4.6's $3, but not the 33% cut the sticker implies. Measure your real token counts before you build a budget on the table. I've watched two clients miss this and wonder why the migration "saved" less than the spreadsheet said.

Prompt caching: the math that decides whether you pay $120 or $66

Caching is the biggest lever on an Anthropic bill and the one most builders leave off. The pricing, as of 2026: a 5-minute cache write costs 1.25x the base input price, a 1-hour write costs 2x, and a cache hit costs 10% of base input (5% on Opus 5.5, 2.5% on Fable 5.1). On Sonnet 5 that's $2.50 per million to write, $0.20 per million to read. On Opus 5.5 it's $5 to write and also $0.20 to read, which makes a cached Opus 5.5 prefix cost the same per token as a cached Sonnet 5 one. If your requests share a stable prefix — a system prompt, a policy document, a product catalog, few-shot examples — every request after the first reads that prefix at a tenth of the price. The write premium pays for itself on the second hit.

What three real workloads actually cost

These are the shapes I see most in owner-led businesses, priced on the September 2026 table. Token counts are the assumptions, so swap in yours.

1. Support bot: 10,000 responses a month, 4,000 tokens in (3,000 of it a stable system prompt and knowledge base), 400 tokens out. That's 40M input, 4M output.

• Sonnet 5, no caching: 40M × $2 + 4M × $10 = $120/month
• Sonnet 5, system prompt cached: 10M uncached × $2 + 30M cached × $0.20 + 4M × $10 = ~$66/month
• Haiku 4.5, no caching: 40M × $1 + 4M × $5 = $60/month (about $33 cached)
• Opus 5.5, no caching: $240/month (about $126 cached; Opus 5 would be $300) — and for a support bot, you won't be able to tell the difference from Sonnet

2. Document extraction: 5,000 invoices or contracts a month, 8,000 tokens in, 500 out, nobody waiting on the result. 40M input, 2.5M output — and "nobody waiting" means Batch.

• Haiku 4.5 standard: $40 + $12.50 = $52.50; on Batch, ~$26
• Sonnet 5 standard: $80 + $25 = $105; on Batch, ~$52.50

This is the workload where model choice is genuinely a cost decision: run a 200-document sample through both, measure the error rate, and let the numbers pick.

3. Content and proposal drafting: 200 long drafts a month, 6,000 tokens in, 2,500 out. 1.2M input, 0.5M output.

• Sonnet 5: $2.40 + $5 = $7.40/month
• Opus 5.5: $4.80 + $10 = $14.80/month (Opus 5 was $18.50)
• Fable 5.1: $12 + $25 = $37/month

At this volume the model is a rounding error against the value of the output. Use the best one. The people agonizing over $11 a month here are the same people paying a writer $2,000 for the drafts.

Anthropic API pricing vs OpenAI API pricing

This page is the Claude side only, on purpose — vendor tables go stale independently, and I'd rather keep one page exactly right than two pages roughly right. The side-by-side, with the OpenAI numbers, the tool-use reliability data from production logs, and the routing pattern most teams end up deploying (Claude for the reasoning hot path, OpenAI for image and voice) is in Claude API vs OpenAI API. The business-level version of the same decision is Anthropic vs OpenAI for business.

If you are pricing the low end of the market too, DeepSeek's API rates, and its weekday peak-hour billing, are on DeepSeek pricing.

OpenAI's own rate card, read from OpenAI's pricing pages, is on OpenAI API pricing. One line on it matters for anyone comparing the two bills: OpenAI bills the API separately from ChatGPT Plus, Business, Enterprise and Edu, so a ChatGPT seat does not cover API calls, the same way a Claude subscription does not cover Anthropic's.

Google is the third quote worth getting, and its rate card has a date on it: Gemini 3.8 Flash is $0.75 in and $3.75 out per million tokens through December 31, 2026, then $1.50 and $7.50 from January 1. Every Gemini rate, and one workload priced against Claude and OpenAI, is on the Gemini API pricing page.

Which Claude model should you actually pay for?

The rule I give every client, as of September 2026: Haiku for volume, Sonnet as the default, Opus 5.5 for judgment, Fable only when Opus 5.5 has already failed the task. Classification, routing, extraction, tagging, anything that runs ten thousand times — Haiku 4.5. The default for everything customer-facing or quality-sensitive — Sonnet 5, now that $2 / $10 is permanent. Contract analysis, complex multi-step agents, the decisions where a wrong answer costs real money — Opus 5.5, which Anthropic's own model guide now names as the place to start. Fable 5.1 is the top of the range and priced like it, at two and a half times Opus 5.5 per token; use it where Opus 5.5 at higher effort still falls short and the volume is small. The longer version, by business task, is which Claude model to use for business.

API pricing vs the subscriptions

Two different meters. The API bills per token, as above. The Claude apps — Pro, Max, Team — bill per month with usage limits instead of a meter, and they're the right buy for humans using Claude directly rather than software calling it. Claude's plan pricing covers that side; Claude Code pricing covers the tool that straddles both, since it runs on either API credits or a Max subscription.

The business-owner answer

The Anthropic API is cheap enough now that the cost question is almost never "can we afford it." It's "which model, cached how, batched when" — and that's a design question, not a procurement one. Get it right and a support bot runs for the price of a lunch; get it wrong and you pay Opus prices for Haiku work with no cache in front of it. If you want the workflow-by-workflow model and cost plan written down for your business, that's the AI Readiness Assessment ($2,500 flat, two weeks, credited against a build within 90 days); the self-serve version is the CFSB Playbook ($397); and the "which of these should we even be running through an API" conversation is free on a strategy call.