Cached context shrinks the discount from Anthropic's half-price Sonnet 5.5

Per token, Sonnet 5.5 costs half of Opus 5.5 on input and output alike: $2 against $4 per million input tokens and $10 against $20 per million output ^[5][6]. Per task, the comparison depends on token count. Sonnet comes out cheaper on a given job only while it burns fewer than twice the tokens Opus would on that job ^[15]. Anthropic's efficiency claim is a cost per task up to 30% lower at an unchanged per-token price, and it is measured against Sonnet 5 ^[3]. The sources do not include a same-task token count against Opus 5.5, and a buyer splitting work between the two needs exactly that ratio.

Cache reads narrow the gap. Both models charge $0.20 per million cached tokens ^[6]. Take an agent job that rereads 10 million cached tokens and writes 100,000 output tokens, leaving fresh input aside. Opus bills $2.00 for the cache and $2.00 for the output, Sonnet bills $2.00 and $1.00, and moving the job down a tier saves 25% ^[16].

Anthropic draws the line between tiers at judgment. It says Opus 5.5 is meant for complex work that needs careful judgment, while Sonnet 5.5 is strongest at well-scoped everyday tasks, bug fixes and polished documents, slides and spreadsheets ^[4]. "Sonnet is really for the cost-conscious customer where they might not need as much intelligence," Theo Chu, a research product manager at Anthropic, told CNBC ^[1]. "It might be routine tasks that just need execution, but don't need that judgment that Opus can bring," he said ^[2]. On GDPval-AA, a test of real-world work across occupations, the model with double the token price scored two points higher ^[10][15].

Anthropic did not cut Sonnet's rate to get there. The $2 and $10 prices were Sonnet 5's too ^[5], so a job that cost $1.00 on Sonnet 5 costs as little as $0.70 on Sonnet 5.5, and all of that saving comes from fewer tokens ^[17]. Early testers reported gains in the same direction. Box said the model was 2.4 times faster and used 12% fewer tokens on sensitive document work ^[13], and Base44 said it needed under half the iterations Sonnet 5 required to finish real app builds ^[14].

The tiering reaches Anthropic's safety spending as well. Because Sonnet 5.5 does not advance the frontier of its models' capabilities, the company said, most of its alignment testing focused on a "targeted set of risks that apply to models of any capability level" ^[11].

The split can land on Anthropic's revenue in more than one way. Opus customers could move routine jobs to Sonnet and pay half per token on them. Sonnet 5.5 could instead mostly absorb Sonnet 5 traffic and work now running on other vendors' models, leaving Opus volume where it was. Haiku 5.5 could pull high-volume jobs lower again once it ships ^[8].

I think the half-price step is nearer a ceiling than a typical saving for teams moving agent work off Opus. Cached context costs the same on both tiers ^[6], and a model two points behind on GDPval-AA may need extra passes on harder jobs ^[10]. The case against: if Sonnet finishes routine jobs in fewer tokens than Opus, the saving exceeds half, and same-task logs showing that would prove this view wrong ^[15].