이미지: METAL LAB 생성
Summary
- Anthropic announced on August 10 that it will keep Claude Sonnet 5's launch pricing in place indefinitely.
- The rate — $2 per million input tokens and $10 per million output tokens — was set at the June launch and had been due to expire on August 31.
- For heavy-token uses like coding agents, the per-token price of a mid-tier model directly determines operating costs.
- 발표 주체·시점
- Anthropic 공식 Claude 계정, 2026년 8월 10일 X 게시
- 대상 모델
- Claude Sonnet 5
- 입력 단가
- 100만 토큰당 2달러
- 출력 단가
- 100만 토큰당 10달러
- 원래 조건
- 6월 출시 시점 도입가, 8월 31일까지 한시 적용
- 변경 내용
- 종료일 없이 동일 단가 유지
A price list that was set to expire on August 31
When Anthropic launched Claude Sonnet 5 in June, it priced the model at $2 per million input tokens and $10 per million output tokens — but only through August 31. With that deadline three weeks away, the company announced via its official account on August 10 that it would keep the pricing in place with no expiration date. The core of the announcement was a single line: "That pricing will remain the same." There was no mention of any change to the model's performance or features.
Why the per-token rate is a number that matters
API pricing is billed by token, not by conversation. A token is the smallest unit into which a model breaks down text, and for English, a million tokens is roughly 750,000 words — the equivalent of several thick novels. The prompts and attached documents a user submits are billed as input, while what the model generates is billed separately as output. It's an industry-wide pattern for output pricing to run about five times higher than input pricing, since generating a response one token at a time is computationally far more expensive than reading through text that's already been provided in one pass.
Where this number really bites isn't for individual users but for the people building services. The coding agents and document-automation tools that have proliferated lately read an entire codebase in a single task and run through dozens of trial-and-error iterations. Tokens pile up into the millions. At that scale, a one-dollar difference in per-token pricing can shift a monthly bill by an order of magnitude. That's why pricing sheets have become something developers check as closely as performance announcements at AI companies.
Launch pricing was supposed to go up
Sonnet occupies the middle tier of Anthropic's model lineup — between the smartest and most expensive Opus and the lighter, cheaper Haiku — making it the workhorse tier most commonly deployed in actual services. This is also the segment where Anthropic collides head-on with OpenAI's and Google's mid-tier models, as well as Chinese and open-weight models offering far lower prices. It has long been a common tactic in the cloud industry to run a limited-time discount on a new model to pull users in, then revert to list price once the window closes. This decision amounts to a declaration that Anthropic won't revert. Once developers build a pipeline around a model, a price increase is enough to get them evaluating alternatives — and Anthropic has chosen not to create that opening.
At the same time, sustaining this pricing requires securing cheap, abundant compute to run inference at scale. That's hard to separate from Anthropic's recent moves to expand its compute base through chip-supply deals and large funding rounds.
AMD to invest up to $7B trillion won in Anthropic, supply 2GW of chips
So what actually changes
Teams building services on Sonnet 5 no longer need to recalculate their pricing after September. For any team that had been budgeting under two different scenarios for what rates might look like after the end of August, that work is now moot. More broadly, this is one scene in a larger trend of frontier-model pricing settling downward at the workhorse tier. Over the past two years, the token cost of achieving the same level of performance has kept falling, and that decline has let automation that used to be demo-only move into actual products. The shift toward coding agents that carry out multiple steps autonomously without human approval has also been made possible by tokens becoming cheap enough to support repeated trial and error. A single quiet line in a pricing sheet still moves the baseline for what kinds of automation are worth leaving switched on.
