Most teams pick a Claude model the way they pick a coffee size: whatever they used last time. That works until the invoice arrives or a response comes back too shallow for the task. Claude now ships in four models (Fable 5, Opus 4.8, Sonnet 5 and Haiku 4.5), and each is tuned for a different kind of work. The skill is to match the model to the job in front of you, rather than reaching for the biggest one out of habit. Getting that match right is one of the cheapest wins available in an AI budget.
Start with the job, not the model
Model selection is really a routing decision. Before you compare benchmarks, describe the job: how much reasoning it needs, how quickly the answer has to come back, how long the output runs, and how much you are willing to spend per thousand calls. Once the job is clear, the model usually picks itself. Teams that skip this step tend to overpay and underperform at the same time. A quick shorthand across the four:
Haiku 4.5: fast and inexpensive, built for high-volume, well-defined tasks.
Sonnet 5: the balanced default for most day-to-day work.
Opus 4.8: the deepest reasoner, for hard problems where quality beats cost.
Fable 5: tuned for long-form and creative drafting where voice matters.
Haiku 4.5: high-volume, low-latency work
Haiku is the model you route to when volume is high and each task is narrow. Classifying support tickets, tagging records, pulling fields out of a form, drafting a first-pass reply, or checking whether a document mentions a topic. These jobs run thousands of times a day, so latency and price matter more than raw reasoning depth. A Sydney logistics firm running 40,000 classification calls a day feels the difference between models directly in its monthly bill. If Haiku handles the task well, sending it to Opus is money spent for no extra accuracy. Reserve the heavier models for the calls that actually reward the extra thinking.
Sonnet 5: the everyday workhorse
Sonnet is the sensible default for work that needs solid reasoning without the top-tier price. Writing and editing internal documents, answering customer questions with nuance, summarising long threads, generating code for common features, and running most agentic workflows. For a lot of Australian small and mid-sized teams, Sonnet handles the large majority of jobs, and you only reach past it when a task genuinely stretches beyond what it returns. The habit worth building is simple: start here, measure the output against what the job needs, and move up or down only when the results tell you to.
Opus 4.8: deep reasoning and hard problems
Opus is for the jobs where a wrong answer is expensive and the reasoning is genuinely hard. Multi-step analysis, tricky refactors across a large codebase, research synthesis, ambiguous decisions with many moving parts, and the planning layer of a complex agent. It costs more per call and can take longer to respond, so you do not want it answering routine questions. You want it on the small share of tasks where the extra depth changes the outcome. A useful pattern is to let Sonnet or Haiku do the routine steps and call Opus only for the hard decision in the middle, so you pay the premium once rather than on every call.
Fable 5: long-form and creative drafting
Fable is tuned for long-form writing and creative work where tone and consistency carry the piece. Blog drafts, marketing copy, narrative documents, and any output where a distinct voice matters more than step-by-step logic. It is a specialist rather than a general reasoner, so pair it with a review step for anything factual. For content teams producing a steady stream of articles, routing the drafting to Fable and the fact-checking to Sonnet is a clean split that keeps quality high and cost sensible. The result reads like your brand rather than a generic template.
A simple routing rule for Australian teams
The cost gap between the smallest and largest model is wide enough that routing pays for itself quickly. A team that blindly sends every request to the top model can easily spend $45,000 a year more than one that routes by job, for output that is no better on the tasks that never needed the depth. There is also a governance angle: under the Privacy Act, keeping a clear record of which model touched which data, and why, makes audits far simpler and gives your board a straight answer when they ask. Build the routing logic once and it keeps returning value. Most teams recover the setup effort within the first month of running at any real volume.
High volume and simple: send it to Haiku.
Balanced everyday work: default to Sonnet.
Hard reasoning or high stakes: escalate to Opus.
Long-form or creative drafting: use Fable, then fact-check with Sonnet.
Unsure: start on Sonnet, measure, then move up or down.
Whatever split you choose, instrument it. Log the model, the task type, the token count and the outcome, then review the numbers monthly. The right routing map for your workload is the one your own data points to, and it will shift as the models improve and your tasks change.
If you want help wiring this routing into your own systems and setting a sensible spend ceiling, we do exactly this with Australian teams every week. You can book a short call to talk it through.
