Haiku 5.5 is a fast, low-cost model for high-volume work such as reading, sorting and pulling details out of documents. It reads images as well as text, handles very long inputs (up to 1 million tokens) and can operate a computer and web browser. It is meant to run alongside Anthropic’s larger models, doing the narrow, repetitive steps of a bigger job.
Two things changed. Price: for requests under 100,000 tokens, Anthropic charges US$0.10 per million input tokens and US$0.50 per million output tokens, against US$1 and US$5 for Haiku 4.5, the previous small model. That is 90% less for those requests and 50% less for longer ones (Anthropic puts the average saving at about 75%), although the new model uses slightly more tokens per task. Ability: on Anthropic’s own computer-use test (OSWorld 2.1, offline subset), Haiku 5.5 scored 72.4% against 15.7% for Haiku 4.5. Those are the vendor’s figures, not an independent test.
For L&V, this means an agent can afford to check every document that comes in, not just the ones a person picks out, and can do simple data entry in web systems that have no integration. A larger model still handles the harder judgement calls, and a person still approves the results.
Checking every carrier invoice line by line against the agreed rate card and booking details, and listing surcharges and fees that don’t match, so the team disputes overcharges instead of paying them because nobody had time to look.
Comparing each new supplier price list against current costs and selling prices, and flagging products whose margin has dropped below target, before the new costs flow through to customer quotes.
Reading site inspection photos and checklists from every visit against each contract’s scope of work, and listing missed areas for supervisors, so problems are fixed before the client notices.
What would it be worth for your team? Enter your team size and hours for a quick estimate.
Try the AI Savings and ROI calculatorOur view: A real change in the cost of running an agent: checks that used to be too expensive to do on every document now cost very little. The computer-use gain looks large but is the vendor’s figure, so we would prove it on a client’s own documents and portals, with a person approving the output, before scaling up.
Tell us about your workflow, your team and the problem you want to solve. We serve Sydney businesses and work with organisations across Australia.