Claude Haiku 5.5: a fast, low-cost model for office work

What Anthropic announced

On 7 October 2026 Anthropic released Claude Haiku 5.5 (opens in a new tab), which it calls ‘the cheapest, fastest, and most capable small model we’ve ever released’. According to the official page, it is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.

On price, Anthropic says that ‘on average, it now costs around 75% less to run’ than Haiku 4.5. A footnote adds that it is priced 90% lower than Haiku 4.5 for requests of up to 100,000 tokens, and 50% lower above that. Anthropic also notes that it uses slightly more tokens per task because of an updated tokeniser, and that this is already reflected in the savings. The official price table lists $0.10 per million input tokens and $0.50 per million output tokens (up to 100,000 tokens), against $1 and $5 for Haiku 4.5.

As recommended uses, the page lists summaries, database queries, classification and live customer support. It adds that Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding.

What it means for your business

A small model fits narrow, repetitive jobs where the right answer is easy to recognise and the volume is high. Three typical examples:

• Classifying: sort emails, support tickets or requests by topic and urgency.
• Summarising: cut minutes, reports or long threads down to five lines.
• Extracting: pull dates, amounts or names out of documents with a similar layout.

For work that needs sustained judgement, long reasoning or decisions with real consequences, use a larger model, and in every case have a person check the results. Anthropic publishes its own comparisons in which Haiku 5.5 scores below Sonnet 5.5: on Terminal-Bench 4.0 it gets 39.2% against 70.6% for Sonnet 5.5. Those are the vendor’s figures, not a guarantee for your use case.

If you process personal data, the GDPR and your company’s internal policies apply as they would with any other model, so find out in which region and under which contract terms the data is processed. If the system talks to customers or members of the public, the EU AI Act requires, among other things, that they are told they are dealing with a machine. A cheap model does not change any of these obligations.

How to try it this week

• Pick a repetitive task that already eats into your time, such as sorting incoming email.
• Prepare 20 real but anonymised examples, with no names, personal data or confidential information.
• Ask the tool to do the task and give it the exact output format (one label, a table, three lines).
• Compare its results with a person’s and note where it gets things wrong.
• Use that result to decide whether automation is worth it, always with a human check before anyone acts on it.

IAescola designs tailored training for teams that want to bring in these tools with good judgement and without unnecessary risk.