What the small Claude is for: Claude Haiku 5.5 in everyday use
What Claude Haiku 5.5 handles well day to day, from email threads to photos of letters, and the moment to move a chat to a bigger Claude.
What is Claude Haiku 5.5 for? It is for the small, steady jobs that fill an ordinary day: boiling down a long email thread, reading a photo of a letter, answering a quick question that needs a look at the web. It does those quickly, and with more care than its size suggests. When a task stops being about finding and condensing and starts to need real judgement, you move the same chat to one of the bigger Claudes, and nothing you have already shared gets lost.
We follow one ordinary afternoon of errands on Haiku 5.5, then show when it is worth stepping up.
A small model that moved up a weight class
Anthropic released Claude Haiku 5.5 on 7 October 2026, and it has been in MultiChats since 9 October. Haiku is the smallest and quickest of Anthropic's Claude lines, with Sonnet and Opus above it. Anthropic calls the new version "our fastest model to date", and says it is the first Haiku with an adjustable effort setting, the control that decides how long a model thinks before it answers.
The bigger change is in what it can handle. On Humanity's Last Exam, a deliberately hard set of expert-level questions across many subjects, Anthropic reports 45.9% for Haiku 5.5 answering without tools, against 10.2% for Haiku 4.5. The same table puts Sonnet 5.5 at 56.9%. Those are the company's own figures from its launch page, so keep them in mind as a claim. The jump is large, putting the small Claude within reach of the middle one on a test built to be hard.
Anthropic pitches Haiku 5.5 at summaries, sorting and live customer support, the kind of work where a fast, dependable answer matters more than a long deliberation. That is most of what people do in a chat window on a weekday.
One ordinary afternoon, start to finish
Picture three small things waiting for you after lunch.
The first is an email thread from your building's residents' group about roof repairs. It is forty messages long, three neighbours disagree, and somewhere in the middle is a question addressed to you. You paste the whole thread into a chat with Haiku 5.5 and ask who has agreed to what, what is still undecided, and whether anyone is waiting on you. This is the work Haiku is built for. It reads the lot and hands back a short account with the open question pulled out at the top. Before acting on a date or sum in the summary, check the message it came from. Summaries capture the shape of a conversation well, but can miss a detail.
The second is a letter. Your energy supplier has written about a change to your tariff, two pages of small print you would rather not read on the sofa while your tea goes cold. You photograph both pages and drop the photos into the same chat. Haiku 5.5 reads images, so you can ask what is changing, from which date, and whether you need to do anything. It will pick out the dates, the new rates and any deadline. Check the figures it quotes against the paper in your hand, because a blurred digit in a photo is the most common way these answers go wrong. Our guide to explaining a screenshot or photo with AI covers which images read well and which quietly fail.
The third is a question that needs today's information: you want to know whether the recycling centre near you is open this Saturday. Haiku 5.5 can search the web, and on the website you usually do not even have to ask for a search. A feature called Intent Detection, switched on for every account by default, reads each message and decides whether it needs one. If you would rather press the button yourself, turn Intent Detection off in Settings, under Tool Preferences, and the search button comes back. The phone apps always show it. The answer arrives with its sources; for opening hours in particular, the venue's own page is the link worth clicking.
How hard is it thinking?
Haiku 5.5 always thinks a little before it answers. In our app there is no option to switch its thinking off; you choose how much, from Low through Medium to High, or leave it on Auto or on the model's own default. For the errands above, the default is fine. Turning effort up helps when a question has several steps that depend on each other, such as working out which of the two payment options in that letter comes out lower for your household over a year.
On the website, the effort control sits beside the search button and follows the same rule. With Intent Detection on, the app sets the effort for each message; with it off, the control is yours. Our piece on reasoning effort explains what each level changes.
When the errand turns into a problem
Now suppose the afternoon goes sideways. The roof thread turns out to hide a dispute over who pays, and the residents' group has a 30-page set of house rules, as a PDF, that says something about it. You want a reply that is fair and faithful to those rules, and that does not start a feud on the landing.
Haiku 5.5 can open that PDF, and it will give you an answer. But now you need it to weigh one clause against another, judge the tone of a message your neighbours will read, keep a dozen details straight across a long document and a long thread, and produce something you will send under your own name.
When your question moves from what does this say to what should I do about it, that is the moment to hand the chat to a bigger Claude.
Step up when the answer has to weigh competing points and reach a judgement, or a mistake could cost you money or a friendship. Long material or a long chat can also be a sign, especially when early details matter. You may also have asked twice and found that the reply still misses the point you care about.
Moving the same chat up a size
Stepping up keeps everything you have done so far. Open the model picker in the same chat and choose Claude Sonnet 5.5 or Claude Opus 5.5. The next reply comes from the model you picked, and it works from the whole conversation: the thread, the photos, the PDF and everything Haiku said along the way. Each reply is labelled with the model that wrote it, so you can always see which Claude said what.
If only one answer needs a second opinion, you can stay where you are. Open the menu on that reply, choose Re-generate and pick a bigger model, and the new answer replaces the old one in place. Try Again does something similar but opens the alternative in a new chat, which is handy when you want to keep both versions.
Which of the bigger two depends on the task, and we compare them in Opus 5.5 vs Sonnet 5.5. If your document is very long, read our look at which AI handles the longest documents first. If a chat has grown so long that the model you switch to cannot hold all of it, the app tells you and offers to switch back or start a new chat. Our piece on long chats explains why very long conversations drift and what to do about it.
Good to know before you pick it
These details were checked on 19 October 2026. Haiku 5.5 needs a Pro or Pro+ plan, and you will find it in the picker's Recommended tab. If you used Haiku 4.5 before, that model is now marked as deprecated in our picker, and choosing it opens Haiku 5.5 instead. Haiku 5.5 reads images and PDFs and can search the web, but it is not one of the models you can talk to in voice mode.
Our view: let the small Claude do the small things
Haiku 5.5 makes a good everyday Claude, and the reasons are speed and fit. Most of what we ask of a chat in a day is reading and condensing: threads, letters, quick facts. Haiku 5.5 does that quickly and well, and by Anthropic's own figures it is a far more capable model than the Haiku of a year ago. When the stakes rise, use a bigger Claude, especially if a reply has to be fair to people you live next to or a long document argues with itself. We would move the chat up on purpose, then move it back once the hard part is done.