Prompt length sets what Anthropic's Haiku 5.5 price cut is worth to each workload

A request-weighted average of Haiku 5.5's two per-token cuts comes to 86% ^[20]. That uses Anthropic's own split: about 90% of Haiku 4.5 requests fell under the 100,000-token line ^[12], where the cut is 90%, against 50% above it ^[3]. Anthropic's published figure is lower, and according to The New Stack it accounts for the request mix and the new tokenizer ^[6]. A long request carries more tokens than a short one, so the long tier weighs more in a token-weighted bill than its tenth of the request count ^[20].

Anthropic flagged the tokenizer change itself and called the increase slight ^[13]. The Decoder notes that the Opus 4.x tokenizer change alone raised token usage by about 30% ^[26]. Taking 30% as the pessimistic case, a short-tier prompt costs the equivalent of $0.13 per million old-tokenizer tokens, 87% under Haiku 4.5 ^[21]. A long-tier prompt costs the equivalent of $0.65, a 35% saving ^[22]. Output lands at the same ratios, because every tier on both models prices output at five times input ^[22].

The Decoder describes the line as a prompt length ^[7]. Neither outlet says whether the higher rate covers the whole request or only the tokens past 100,000. If it covers the whole request, a 101,000-token prompt costs about $0.05 in input against about $0.01 for a 99,000-token one ^[23]. A prompt that sat just under the line on Haiku 4.5's tokenizer can cross it on Haiku 5.5's ^[13]. Compaction, one of the jobs Anthropic recommends the model for, tends to start from a long context ^[14].

Haiku 5.5 is also the first Haiku with effort controls, and the setting governs how many tokens it spends on a task; the default is medium ^[9]. A per-task cost measured at medium is the figure to set against Haiku 4.5 ^[9]. Raising effort to recover accuracy on hard tickets is paid for in tokens ^[9].

The quality case rests on Anthropic's own evaluations ^[10]. Haiku 5.5 scored 72.4% on the offline OSWorld 2.1 subset, against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna ^[10]. On Terminal-Bench 4.0 it went from zero to 39.2%, a gain best not quoted as a multiple, while Sonnet 5.5 scored 70.6% ^[11]. Both suites test computer use and command-line work, and neither scores label accuracy on a support queue ^[10][11]. I would not move a classification workload on these tables without first scoring a labelled sample of its own traffic at medium effort. Anthropic's guidance keeps Haiku 5.5 on narrowly scoped work and leaves complex agentic coding to Sonnet 5.5 and Opus 5.5 ^[14].

Anthropic's 20% estimate for Sonnet 5.5 implies that cache reads were about 40% of a typical agentic task's cost before the cut ^[24]. A loop with a smaller cached share saves less ^[24]. Haiku 5.5's short-tier input and Sonnet 5.5's cache reads now cost the same $0.10 per million tokens ^[25]. According to The New Stack, the Sonnet cut rolls out across platforms on launch day, though some existing Azure and Google Cloud customers will wait a few days ^[16].