AR Digital / Daily AI
Opus, Sonnet or Haiku: pick the right Claude model to protect your usage limits
Sonnet by default, Haiku for simple tasks, Opus for complex ones: the rule for picking the right Claude model and protecting your usage limits mid-day.

Radu UleaFounder of AR Digital

Stopped mid-day by the “usage limit reached” message? In the teams we train, the limit often goes less on big analyses than on simple tasks handed to the most powerful model. Choosing the model to suit the task makes a real difference.
Daily AI. Today’s tip: pick the right Claude model to protect your usage limits. Claude offers several models, and they don’t use up your limit at the same rate. Use Sonnet by default, for writing and analysis. Switch to Haiku for simple tasks: summarising, finding a piece of information, sorting emails. And keep Opus for genuinely complex problems, when Sonnet has shown its limits. Two extra habits: open a new conversation when you change subject, and store your recurring documents in a project. And before each request, one simple question: do I want to find, write or reason? The detail is just below.
Claude offers several models, and they don’t use up your usage limit at the same rate. Use Sonnet by default, Haiku for simple, quick tasks, and keep Opus for complex problems where Sonnet has shown its limits. You can work for longer before hitting the limit.
The problem: a usage limit that melts away for no reason
Claude’s Pro, Max and Team subscriptions don’t give you an unlimited number of exchanges. On the Pro plan, the session limit resets every five hours, and a weekly limit applies to all models, according to the Pro plan help page. The best practices published by Anthropic list what makes consumption vary: the length of messages and attached files, the length of the current conversation, tools such as web search, the level of effort requested and the choice of model.
That last point is the one many users overlook. Many pick the most powerful model once and for all, thinking “whoever can do more can do less”. For quality, that is sometimes true. For your usage limit, it is the opposite: every simple request sent to a heavy model such as Opus costs far more than it should.
The habit: one model per type of task
Anthropic describes four models in its guide to choosing the right model: Haiku, Sonnet, Opus and Fable. Fable, offered with paid plans, is aimed at very large projects and uses the most of your limit. For an SME’s everyday work, the first three are the ones that matter. Remember one simple rule: Sonnet by default, Haiku when it’s simple, Opus when it’s genuinely hard.
- Haiku, for instant answers. Quick replies, summaries, simple extraction of information. The guide presents it as the model that is gentlest on your usage limit.
- Sonnet, for everyday work. Writing, analysis, code and multi-step tasks. Anthropic makes it the default all-round model, and advises starting with it when in doubt.
- Opus, for long reasoning. In-depth research and complex problems that call for sustained thinking. The guide warns that it uses more of your limit and should be kept for tasks that really need it, for example those where Sonnet has already struggled.
For Claude Code, Anthropic’s tool that makes Claude work as an agent, the help page on models and limits is even more direct: Opus costs several times more per exchange than Sonnet, and Sonnet more than Haiku. Using Opus for routine work is, according to Anthropic, the fastest way to exhaust a limit.
The principle isn’t limited to Claude. ChatGPT and Gemini also offer several models of different power, with limits that depend on your plan: your subscription’s help page will tell you which limits apply to each model.
A step-by-step example
Let’s take a fictional six-person estate agency, subscribed to Claude, on an ordinary day.
- In the morning, sorting emails. Summarising twenty viewing requests and pulling out the dates people want: that is simple extraction. Haiku is enough.
- Late morning, writing the listings. Three property descriptions from the viewing notes, in the agency’s tone: that is routine writing. Sonnet is the right choice.
- At the end of the day, analysing a thorny file. Comparing a complex sales agreement with the parties’ correspondence to spot inconsistencies: Sonnet first. If its answers stay superficial, the same request moves to Opus. In every case, a competent person reviews the analysis before any decision.
The question to ask before every request. “Am I asking it to find, to write or to reason?” Finding information in a text and summarising: Haiku. Writing and analysing: Sonnet. Reasoning at length on a hard problem: Opus, after a try on Sonnet.
Anthropic’s guide also suggests spending a few minutes testing your usual tasks on the different models. Once the team knows what each one does well, the choice becomes a reflex.
Pitfalls to avoid
The first pitfall is leaving Opus selected for everything, out of habit. Check the model shown before you write, especially when opening a new conversation.
The second pitfall is the endless conversation. Because the length of the current conversation counts towards consumption, a thread that has been dragging on since the morning weighs down every new question. When you change subject, open a new conversation. Anthropic’s best practices add three useful tips: group similar requests in a single message, refer to what has already been said rather than repeating it, and store recurring documents in a project, whose cached content counts for less against your limits when it is reused. That cache expires after a period of inactivity: when you come back, the first message counts that content in full again.
That last point ties in with our tip on the context sheet: once stored in a project, it saves you pasting the same information into every request.
Our take
In the teams we train, most of the time, hitting the limit is not a sign that you need a more expensive subscription. It is a sign that you are using a bulldozer to plant nails. Learning to choose the model takes little time, and the gain shows quickly: less waiting, less frustration, and the powerful tool available on the day it is really needed.
Taking it further with your team
The same logic applies when you hand repetitive tasks to an AI agent. An agent that sorts, files or summarises hundreds of items a day doesn’t need the most powerful model: that is precisely Haiku’s territory, according to Anthropic. Our own agent runs on a local model installed on a single Mac in the office: we describe how we put it to work in our Hermes feedback article.
In Luxembourg, training on these uses can be partly funded as part of an AI integration project: we explain how in our guide to funding AI training through SME Packages.
Frequently asked questions
Which Claude model should I choose when I’m unsure?
Sonnet. Anthropic presents it as the default all-round model and advises starting with it when in doubt. Switch to Haiku for simple tasks, and to Opus only if Sonnet struggles with a complex problem.
Does Opus always give better answers?
Not on simple tasks. According to Anthropic’s guide, using Opus for a task that Haiku could handle uses up your allowance with no gain and slows the work down. Opus is justified for problems that call for long thinking.
When does my usage limit reset?
On the Pro plan, the session limit resets every five hours, and a weekly limit applies to all models, according to Claude’s help. The time of the next reset is shown in Settings, under Usage. The detail depends on your plan.
Does the tip still hold for ChatGPT?
The principle does: ChatGPT also offers several models, more or less powerful, with limits that depend on the plan. Names and caps differ, so check your subscription’s help page before setting your team’s rules.
Every working dayAR Digital / Daily AI
Get Daily AI every working day.
One piece of AI news, what it changes for your business and what to do, in your inbox in the morning. Free, no commitment.




