Task statement 3.3 · Product and model selection
3.3 · Matching Models to Tasks
In Wrenfield Housing, the course example, two jobs land in the same week. The operations team must draft short appointment confirmations for the 112 damp and mould jobs, and the communications lead must draft the board's case for a second contractor team. One model choice for both jobs wastes time and usage on the simple job, or quality on the hard one.
share
The tempting habit is to pick the most capable model for everything, because it feels safe. On simple, repetitive work it answers more slowly and uses up your usage limit faster, and a 50-word confirmation gains nothing a reader would notice. The opposite habit fails too: one fast model for everything saves almost nothing on a single paper, while a weak argument in front of the board costs far more than the minutes saved.
KnowledgeStart from what the task needs
As 3.2 showed, you pick the model from the menu next to the send button and can change it mid-conversation. Haiku is the fastest and lowest-cost family, Opus the most capable of the guide's three, and Sonnet sits between them. Anthropic's guide to choosing a model says a faster, more cost-effective model suits high-volume, straightforward tasks. The most capable model suits complex reasoning and work where accuracy outweighs cost. So the choice starts from the task, which rules out picking one model for every job. Describe the task first:
- How hard is the reasoning? Filling known facts into a fixed pattern is easy; weighing options against each other is hard.
- How much volume? One paper or a hundred messages: every extra second and bit of usage is multiplied.
- How fast is an answer needed? A person waiting on each reply feels a slower model.
- What does a mistake cost? A slip in a short message is easy to spot in a check; a gap in a long argument is hard to spot and can sway a decision.
Interactive · one model choice, two jobs
Choose a model for both jobs and see where cost, speed and quality break.
Most capable for both: the confirmations pass, but the officer reaches the session limit at batch 9, so the board case's figures wait five hours.
Fastest for both: the confirmations pass quickly, but the board draft weighs no risks or options and passes 2 of 4 checks.
Matched: Haiku for the confirmations, Opus with thinking on for the board case. Both pass, and nobody waits for a reset.
KnowledgeWhat cost, speed and quality mean in the apps
In the Claude apps, cost is rarely a price per message. It means your usage limit, your time, and money only when your organisation has turned on usage credits. A usage limit is the amount of Claude use a plan allows over a period. On a Team plan the limits are per member, with a session limit that resets every five hours and a weekly limit across all models. The Claude Help Center says the model and the effort level you choose affect how fast you use your limit. Its article on usage credits recommends Haiku or the newest Sonnet model for most tasks, at lower token costs than Opus models. A member who reaches a limit waits for the reset, unless the organisation has enabled usage credits.
Speed is the wait for each reply multiplied by the number of replies, so it matters for 14 batches and hardly at all for one paper. Quality means passing the checks the task's readers set, not the best text any model could write. A more capable model adds quality only where the task needs reasoning a faster model gets wrong, which rules out "the most capable model is always safe". The two ends compare like this:
Faster, lower-cost model
Short replies from known facts, many times over, where a slip is easy to spot. Haiku, in the guide.
Gains: quick replies, and your usage limit lasts the whole job.
Most capable model
Weighing options or building an argument for the board, once or a few times. Opus, in the guide.
Gains: reasoning that holds up, at the price of slower replies and more usage.
Two more settings in the same menu trade quality against speed and usage inside one model. The effort level controls how much thinking Claude applies to a response: higher effort is more thorough, but slower and heavier on your limit. Extended thinking is a setting, shown as a Thinking or Extended toggle, that lets Claude spend more time breaking down a problem before it answers. The Help Center advises raising effort, turning on thinking, or both for complex tasks. Simple tasks need neither, and lower effort makes your usage go further. On some of the newest models thinking cannot be turned off.
SkillWorked example: 112 confirmations on the wrong model
The contractor books visits for Wrenfield's 112 damp and mould jobs each week, and the operations officer drafts the confirmations in batches of eight. Each prompt holds only the job reference, estate, date and slot, because AI policy rule 1 keeps tenants' details out. The repairs system adds names and addresses when it sends. The same officer owes the communications lead the board case's figures by 14:00:
# before: Opus family, thinking on, for all 14 batches Draft a confirmation for each job: reference, estate, date, slot. → about 1 minute a batch; session limit reached at batch 9 (11:40) → figures section for the board case waits 5 hours for the reset # test: 5 real jobs on the Haiku family, same prompt and checks check: reference · date · slot · rebook via the repairs line · 60 words or fewer → 5 of 5 pass, about 10 seconds a batch # after: confirmations on Haiku; board case on Opus, thinking on → 14 batches before 10:00; board draft passes 4 of 4 board checks
The symptom was a missed deadline, not a bad draft, and the cause was a model chosen by habit. Filling four facts into a fixed message is easy reasoning at high volume, so the most capable model spent a minute a batch and most of the officer's session on quality the checks could not see. Because the job recurs every week, the officer did not switch blind: five real jobs went through the faster model against the same checks first. The board case went the other way. It must weigh £48,000 against cutting the median wait from 41 days to 20, and say what happens if the contractor's survey slips past 28 November, so it stayed on Opus with thinking on. The officer still checks each batch before it is sent, and the operations director checks the board paper, as Wrenfield's AI policy requires.
| Simple, repetitive, high volume | A faster, lower-cost model, after a few real examples pass your checks. |
| Complex reasoning or high stakes | The most capable model, with thinking or higher effort on. |
| Mixed, or you are not sure | Test a few real examples on the middle family, then move up or down. |
- In the Claude apps, cost means your usage limit and your time, and money only when usage credits are turned on.
- The most capable model is no safe default: on simple, repetitive work it spends limit and time on quality nobody sees.
- Before moving a recurring job to a faster model, run a few real examples through the checks you already use.