Anthropic makes Sonnet 5's launch pricing permanent, scraps September increase
The scheduled step-up to $3 and $15 per million tokens is off. For companies running Sonnet 5 agents in production, Sunday's quiet announcement removes a 50 percent cost cliff.
Anthropic will keep Claude Sonnet 5's introductory API pricing in place permanently, the company said Sunday, canceling a scheduled September 1 increase that would have raised the model's rates by 50 percent.
Sonnet 5 launched on June 30 at $2 per million input tokens and $10 per million output tokens, billed at the time as introductory pricing good through August 31. The standard rate was set to move to $3 and $15 next month. That step-up is now dead. Anthropic's launch announcement has been updated to reflect the permanent rates, and the reversal was picked up in a Techmeme roundup within hours.
Why the cliff mattered
For most API customers a price change is an annoyance. For agent operators it is a margin event. Sonnet 5 was pitched squarely at agentic work from day one. TechCrunch's launch coverage described a model that can make plans, use tools like browsers and terminals, and run autonomously, at rates well under Opus-class pricing. Companies reselling that capability price their products per task, per resolution, or per seat, and those prices get set weeks or months ahead. A 50 percent jump on both the input and output legs would have landed directly on gross margin for any operator unable to reprice by September 1.
Output tokens are the sharper edge. Agent workloads skew output-heavy because plans, tool calls, retries, and self-checks all count as generation. Holding output at $10 rather than $15 per million tokens is the difference between a viable unit cost and an awkward conversation with customers for services that run thousands of agent loops a day.
A price war artifact
The decision reads less like generosity than like a market position. Sonnet 5's launch pricing was itself a move in the summer's agent-model price war, the mid-tier scrap we charted in July's pricing roundup, and Google has since cut Gemini 3.6 Flash rates as covered in our earlier report. Raising the mid-tier in September would have made Anthropic the only frontier lab moving prices upward this year, an uncomfortable place to stand while courting agent builders who compare per-task costs in spreadsheets.
The announcement covers Sonnet 5 only. Anthropic said nothing about Opus 5, which it already repositioned on cost and capability this summer, or about extended-context rates for its other models. Operators blending Sonnet 5 execution with Opus 5 planning still carry two separate pricing exposures, and only one of them just got locked down.
What to check this week
Two practical items follow. First, any GaaS contract indexed to model list prices, a clause some vendors added during the spring's volatility, no longer needs a September rider for Sonnet 5 workloads. Second, capacity planning that assumed a migration to a cheaper model on September 1 can be shelved, along with the evaluation work that usually accompanies a forced model swap. Permanent, of course, means permanent until the next announcement. But in a market where prices have mostly moved in one direction since June, a frozen price is the closest thing agent operators get to certainty.