Anthropic lowers small-model API prices with Haiku release

A lower-cost model with an effort control

Anthropic has announced Claude Haiku 5.5, targeting high-volume tasks including summarisation, context compaction, database queries, classification and subagents. The October 7 release is the first Haiku model to offer adjustable reasoning effort.

That control lets developers vary the amount of processing assigned to a task. It adds a capability-and-cost decision beyond choosing the model alone, although the announcement’s performance and savings estimates remain company assessments rather than independent Quickanomics measurements.

The model is listed as claude-haiku-5-5 through Anthropic’s API, AWS, Google Cloud and Microsoft Azure. Availability and the price table concern the released product; they should be kept separate from the vendor’s evaluation of how it performs on a particular workload.

The pricing comparison

For contexts up to 100,000 tokens, Haiku 5.5 costs US$0.10 per million input tokens and US$0.50 per million output tokens. Cache reads cost US$0.01 and cache writes US$0.125 per million tokens. Above the context threshold, the corresponding rates are US$0.50, US$2.50, US$0.05 and US$0.625.

Haiku 4.5’s listed rates were US$1 for input, US$5 for output, US$0.10 for cache reads and US$1.25 for cache writes. The new rates represent nominal unit-price cuts of 90 per cent below the threshold and 50 per cent above it.

The new tokenizer complicates the comparison because the same work can produce more tokens. Anthropic estimates average cost roughly 75 per cent lower after accounting for tokenisation, and says about 90 per cent of earlier requests fell below the context threshold. A customer’s actual bill still depends on context size, generated output, cache use and effort settings.

The announcement also lowers Sonnet 5.5’s cache-read rate from US$0.20 to US$0.10 per million tokens, with an estimated saving of about 20 per cent for a typical agentic workflow. Monthly extra-usage credits are described as up to US$100 for Max 5x, US$200 for Max 20x and a pooled US$500 for Teams. Rollout and eligibility conditions apply; the Teams figure is a pool rather than an entitlement of that size for each member.

Reported performance and safety

Anthropic reports an offline OSWorld score of 72.4 for Haiku 5.5, against 15.7 for Haiku 4.5. Its Terminal-Bench 4.0 result is 39.2, compared with 70.6 for Sonnet 5.5. Humanity’s Last Exam scores are 45.9 without tools and 57.4 with tools.

These measures concern different evaluations and tool settings. They indicate the vendor’s reported results, rather than a guarantee that the same improvement will appear in every production task.

The company’s safety account places Haiku 5.5’s cybersecurity capability above Haiku 4.5 and below Sonnet 5.5, with biology capability similar to Sonnet 5, Sonnet 5.5 and Opus 5. Those findings should remain attributed to Anthropic and should not be converted into a claim of zero risk.