Claude Sonnet 5 pricing: the September increase is not happening

Claude Sonnet 5 pricing: the September increase is not happening

Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens, with a 1,000,000-token context window. That price was announced as introductory through 31 August 2026, but Anthropic has made it standard: the increase to $3/$15 set for 1 September will not happen.

Data last verified: 29 August 2026, against the official Anthropic documentation linked at the end.

If you looked this up a few weeks ago, you probably read the opposite. For months the public information said Claude Sonnet 5 was going up 50% on 1 September, and plenty of articles still say so. It is no longer true, and it is worth knowing before you sign anything justified by that increase.

Is Claude Sonnet 5 pricing going up on 1 September 2026?

No. Anthropic's pricing documentation states it plainly:

“The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.”

The practical takeaway: there is nothing to do on 1 September. No migration, no renegotiation, no pulling usage forward.

What exactly does Claude Sonnet 5 cost?

Input price$2 per million tokens
Output price$10 per million tokens
Context window1,000,000 tokens
Maximum output128,000 tokens
API identifierclaude-sonnet-5
Cache read$0.20 per million tokens
Cache write (5 minutes)$2.50 per million tokens
Batch processing50% discount on input and output
Price scheduled for 1 September 2026Cancelled; stays at $2/$10
Committed retirementNot sooner than 30 June 2027

Two details that rarely make the comparisons and that move the bill more than the list price. First: cache reads cost $0.20 per million tokens, a tenth of the input price. If your agent repeats the same long context on every conversation, caching is the difference between paying for that text once and paying for it five hundred times. Second: if the work is not time-sensitive, batch processing halves the price.

What does an AI agent on Claude Sonnet 5 cost per month for a small business?

This is where dollars per million tokens stop being an abstraction. Take a concrete, ordinary case: a customer-service agent handling 500 conversations a month. Assume 3,000 input tokens and 700 output tokens per conversation — about 3,700 in total, the average Anthropic's own documentation uses in its customer-support example — with no caching and no batching, which is the worst case.

ModelModel costEuro equivalent
Claude Haiku 4.5$3.25 per month€2.78 per month
Claude Sonnet 5$6.50 per month€5.56 per month
Claude Sonnet 4.6$9.75 per month€8.33 per month
Claude Opus 5$16.25 per month€13.89 per month
Claude Fable 5$32.50 per month€27.78 per month

European Central Bank reference rate for 21 August 2026: €1 = $1.1699. Rates move; dollars are what Anthropic actually bills.

Read that twice, because it breaks the usual conversation: that agent costs $6.50 a month in model spend. Had the increase gone ahead, it would have been $9.75 — $3.25 more per month. A 50% rise on a number that moves nobody's budget.

The uncomfortable conclusion for anyone selling AI on the strength of which model they use: the model is not what costs money. What costs money is integrating it with your CRM, maintaining it when the API changes, making sure it does not hallucinate in front of a customer, and having someone answer when something breaks on a Friday. If an AI agent quote is justified mainly by the model price, it is badly explained.

What if your provider raised prices “because of the September increase”?

That is a fair reason to ask for a review. The increase was announced until recently, so anyone who passed it through was not acting in bad faith — they worked with the information available. But it no longer exists, and a rate justified by it should come back down.

What you can ask, without being technical: which model exactly is used — the precise identifier, not “the latest Claude” — how many tokens your case consumes per month, and which part of the invoice is model and which part is service. A provider who cannot answer that in an email has a transparency problem, not a pricing one.

Where is your data processed? The detail the comparisons leave out

If you work in a law firm, a clinic, or anywhere GDPR is not a formality, this matters more than price. Anthropic's data residency documentation is explicit about the available options:

“Inference geo: Only "us" and "global" are available. Workspace geo: Only "us" is currently available.”

There is no European geography to pin to. Either you accept global routing, which is the default and lets inference run in any available region, or you pin the United States at a 10% premium. It is official, checkable, and we have not seen it covered in Spanish in any pricing comparison.

This does not mean you cannot use these models under GDPR. It means the international transfer belongs in your record of processing activities and your assessment, rather than being discovered in an audit. And if your case genuinely requires that data does not leave, the conversation stops being about which model is best and becomes one about architecture: what gets sent to the model, what is anonymised first, and what never leaves at all.

What we do with all this

At AizuaLabs we build AI agents for small and mid-sized companies, and we pick the model case by case, not by headline: the cheap one to classify and route, the mid one to converse, the expensive one only where it shows. We bill a flat price, with no credits and no end-of-month surprises, precisely because — as you have just seen — the model share is the small, predictable part.

If you want to know what your case would cost, with your volumes and your data in front of us, the quick way is a free 60-minute audit: you leave with the numbers written down, whether or not you work with us. You can also look first at how we build AI agents for companies.

Frequently asked questions

How much does Claude Sonnet 5 cost per million tokens?

Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens on the Claude API. Cache reads cost $0.20 per million and batch processing applies a 50% discount.

Is Claude Sonnet 5 pricing increasing on 1 September 2026?

No. Anthropic cancelled that increase and made the $2/$10 introductory pricing the standard price. The rise to $3/$15 scheduled for 1 September 2026 will not happen.

How large is the Claude Sonnet 5 context window?

Claude Sonnet 5 has a 1,000,000-token context window and can generate up to 128,000 output tokens in a single response.

Which Claude model is the cheapest?

Of the current family, Claude Haiku 4.5, at $1 per million input tokens and $5 output. In exchange, its context window is 200,000 tokens, not one million.

What does a customer-service agent on Claude Sonnet 5 cost per month?

For 500 conversations a month at about 3,700 tokens each, model spend on Claude Sonnet 5 is roughly $6.50 per month (€5.56 at the ECB rate for 21 August 2026). The real cost of a production agent is dominated by integration and maintenance, not by the model.

Can Claude data be processed inside the European Union?

On the Claude API there is no European geography to pin to: the only options are global routing, which is the default, and the United States, at a 10% premium. The international transfer must be documented in your record of processing activities.

How long will Claude Sonnet 5 be available?

Anthropic commits to not retiring it before 30 June 2027 on its own platforms. Amazon Bedrock and Google Cloud set their own schedules.

Sources

We revise this article whenever one of those figures changes. If you spot an out-of-date number, tell us and we will correct it.