How much does Claude Sonnet 5 cost, and why the price is not going up
Anthropic cancelled the price increase for Claude Sonnet 5 that was scheduled for September 1, 2026. A plain-language look at the API rates: what it costs and who it affects.
Were you about to build something on Claude and bracing for a price hike on September 1, 2026? It is not happening. Anthropic, the company behind Claude, has made the launch price of Claude Sonnet 5 permanent. Here is what that costs and who it actually affects.
Which price is this: the subscription or the API?
First, a fork so nothing gets confused. Claude has two separate ways to pay.
One is the subscription on claude.ai (the Pro and Max plans). You pay a fixed amount per month and talk to Claude in the chat, browser, or app. You never see tokens and the bill does not grow with each message. This article is not about that.
The other is the API (the interface your own programs use to call Claude directly). Here you pay by volume: for every chunk of text you send the model and get back. That chunk is called a token. The moment you build something of your own on Claude, an automation, a support agent, an integration through MCP (the Model Context Protocol, which connects Claude to outside services), or you write code in Claude Code, the bill runs through the API and is counted in tokens. That is the price this news is about.
What changed, or rather did not?
At the launch of Claude Sonnet 5, Anthropic set an introductory price: $2 per million input tokens and $10 per million output tokens. It was billed as temporary, running through August 31, 2026, after which the price would rise to $3 and $15 on September 1.
On August 10, 2026, Anthropic cancelled that increase. The $2 and $10 rate is now the standard price, not a promotion. In Anthropic's own words on the pricing page, the previously scheduled increase to $3/$15 per million tokens on September 1, 2026 will not occur. You can check it in the release notes and on the pricing page.
A million tokens (written MTok in the tables, for million tokens) is a lot. One token in English is roughly 0.75 of a word, by Anthropic's own estimate. A million tokens runs to hundreds of pages.
What does that cost in real money?
Take a plain example. Say you run two hundred customer messages through Sonnet 5 so the model sorts them and drafts replies. Assume each message is about 1,000 tokens in and 500 tokens of reply.
Input: 200 times 1,000 is 200,000 tokens, or 0.2 million. At $2 per million, that is 40 cents.
Output: 200 times 500 is 100,000 tokens, or 0.1 million. At $10 per million, that is $1.
Total: $1.40 to process two hundred messages. At the would-be $3 and $15 rate, the same job would have cost $2.10. Cancelling the hike saves about a third on every task like this. On a one-off it is pennies; across thousands of operations a month, the gap adds up.
How else can you shrink the API bill?
Three levers, all from the official pricing page.
Batch processing. If a job is not urgent and you can wait up to a day for the result, switch to batch mode and pay half. For Sonnet 5 that is $1 per million input and $5 per million output.
Prompt caching. If you send the same large block of text, an instruction or a knowledge base, with every request, you can cache it. A cache read costs 10 percent of the normal input price.
The right model for the job. Sonnet 5 is the workhorse for most tasks. For very simple operations, Haiku 4.5 is cheaper ($1 in, $5 out). For the hardest reasoning, Opus 5 costs more ($5 in, $25 out). Do not reach for Opus where Sonnet is enough.
Who does this matter to right now?
If you only use Claude in the chat through a subscription, nothing changed for you and there are no tokens to count. The news is for people who build: automations, agents, integrations, anyone writing in Claude Code. For them Sonnet 5 stays one of the best-value models at its level, and now for the long run, not just through the end of August.
The fastest way to see where Claude actually saves you time is to try it. We put together free materials: guides and a first hands-on lesson at /try/b0-01-unit, where you build your first working result in about fifteen minutes.
AGINE Academy is an independent product and is not affiliated with Anthropic. Claude is owned by Anthropic.
Questions
No. Anthropic cancelled the planned increase on August 10, 2026. The rate of $2 per million input tokens and $10 per million output tokens is now permanent.
No. The Pro and Max plans are a fixed monthly fee and do not count tokens. This news is only about API billing, which applies when you build your own programs on Claude.
Add up the rough input and output token counts, divide by a million, and multiply by $2 and $10. For non-urgent jobs, switch on batch mode and halve the price.