Claude's Cheap Agent Tokens Expire on September 1
Introductory rates on Anthropic's mid-tier model run out at the end of August, lifting the per-token bill by half. Teams that budgeted during the promotion have four weeks to rework the maths.
Monday, August 3, 2026/2 min read
Introductory pricing is a marketing decision that behaves like a deadline. Anthropic set one when it launched Claude Sonnet 5, and it lands in four weeks.
From 1 September, the model moves from two dollars per million input tokens and ten per million output to three and fifteen. For anyone running it at volume, that is a flat fifty percent increase arriving on a single date.
What changes on September 1
The promotional rate was published as a fixed term rather than an open offer. Anthropic's launch announcement set the introductory price through 31 August, with standard pricing taking effect the following day, and the current rates remain listed on the Claude platform pricing page.
Nothing about the model changes. Only the invoice does.
Why agent workloads feel it hardest
A chat assistant consumes tokens in bursts shaped by human typing speed. An agent does not. It reads files, plans, calls tools, checks its own work and loops, and every one of those steps is billed.
Anthropic raised rate limits across Chat, Cowork, Claude Code and the platform precisely because higher effort levels burn more tokens. That is a fair accommodation, and it also means the workloads the company has been encouraging are the ones most exposed to a fifty percent rise.
An independent breakdown from FinOps LLM puts the same point in blunter terms: teams that sized their budgets during the introductory window are looking at a step change rather than a drift.
The other August deadlines
The pricing cliff is not arriving alone. A running timeline of Claude usage limit changes records a temporary fifty percent weekly usage boost for Claude Code subscribers extended through 19 August, and the developer platform is retiring its legacy Workbench and experimental prompt tools APIs, with access ending 17 August.
Read together, August is a month of expiries. A usage boost ends, two legacy interfaces close, and a price returns to its standard level.
How teams are responding
The obvious lever is tier selection. Work that genuinely needs a stronger model keeps it, and everything else moves down. The less obvious lever is effort control, since a model that can be told how hard to think is also a model that can be told when not to.
Prompt caching and shorter context windows do the unglamorous rest. None of this is new advice. What is new is that ignoring it now carries a dated cost.
What to watch next
Introductory pricing tells you what a vendor wants you to try. Standard pricing tells you what the vendor thinks the work is worth. The gap between those two numbers, and whether customers stay once it closes, is the honest measure of how much of the current agent boom is genuine demand and how much of it was a discount.
Published in The Outspoken Digest
Newsletter
The Digest, in your inbox
One edition, sent when it is ready. No noise, and your address is never passed on.
Read Next
More Technology →
