Haiku 5.5 raises a practical question for teams with frequent summaries, classification requests, or short information queries: when does a smaller model become a sensible way to reduce spending? The strongest case is a narrow task with a clear result. The price announcement makes that case attractive, but it does not establish that every workload will receive the same savings.
In its October 7 announcement, Anthropic positions Claude Haiku 5.5 for high-volume, cost-sensitive work, including summarization, compaction, queries, and classification. Those examples suggest a useful starting point: examine the repetitive work that does not need the full capabilities of a larger model. This is a recommendation based on the vendor’s stated positioning, rather than a conclusion from independent testing.
The published prices have two prompt-length bands. For prompts of up to 100,000 tokens, Anthropic lists input at $0.10 and output at $0.50 per million tokens. Above 100,000 prompt tokens, the listed rates are $0.50 for input and $2.50 for output. The same table lists Haiku 4.5 at $1.00 for input and $5.00 for output per million tokens. At an unchanged token count, that comparison means 90% lower token prices in the shorter-prompt band and 50% lower prices in the longer-prompt band.
That distinction matters before any discussion of an average. A workload dominated by short requests has a different pricing starting point from one built around very large prompts. Output also remains five times as expensive as input per token in either Haiku 5.5 band. Consequently, a concise classification answer and a lengthy generated summary can have different cost profiles even when they start from the same amount of input.
Anthropic also lists separate cache rates. In the shorter-prompt band, cache reads cost $0.01 and cache writes $0.125 per million tokens; in the longer-prompt band, they cost $0.05 and $0.625. These are another reason to keep the workload’s input, output, and cache mix visible when considering the price change. A single headline percentage cannot describe every combination.
The vendor’s larger claim is that Haiku 5.5 reduces the cost of completed work by roughly 75% on average compared with Haiku 4.5. That is a vendor-reported task-cost estimate, with conditions. Anthropic says 90% of Haiku 4.5 requests fell within the lower prompt band, and it notes that the updated tokenizer produces slightly more tokens for a task. The average therefore should not be treated as a flat billing discount or a guaranteed forecast for another team’s workload.
Haiku 5.5 also introduces adjustable effort to the Haiku family. Anthropic presents this as a way to choose between cost and intelligence; the announcement does not establish one universally optimal setting. The practical recommendation is to define what a task must accomplish before choosing how much effort it warrants. A lower price is useful when the resulting answer still meets that requirement.
The scope of the decision matters as much as the rates. Anthropic continues to recommend Sonnet 5.5 and Opus 5.5 for complex agentic coding. For readers considering Haiku, the useful takeaway is to start with frequent, bounded tasks and budget for their actual prompt lengths and output needs. The launch provides compelling published rates for that choice. Its average savings claim remains context for a decision, rather than a substitute for understanding the work being purchased.