Task statement 3.2 · Product and model selection
3.2 · Claude Model Families
In Wrenfield Housing, the course example, two officers in the operations team sent Claude the same prompt, the message that asks for the work, and got noticeably different replies. One of them had switched the model selector to a different model family for an earlier job and forgotten, so the gap came from the model, not from the wording.
share
When two people get different replies to the same prompt, the natural move is to blame the wording and start rewriting it. That cannot help when the replies came from different model families, because each family is built for different work and answers at its own speed and depth. The opposite mistake is to treat the most capable family as right for everything, which ignores that it is also the slowest and costs the most.
KnowledgeThree families, one trade-off
A model is the trained system that writes Claude's replies, and Anthropic releases its models in families. The exam guide names three: Haiku, Sonnet and Opus. The families trade capability against speed and cost. Opus is the most capable of the three and the slowest. Haiku is the fastest and lowest-cost. Sonnet sits between them, and Anthropic's models overview calls it the best combination of speed and intelligence. Anthropic's price list charges per token, a small piece of text that models read and are billed in. Opus costs more per token than Sonnet, and Sonnet costs more than Haiku. In the Claude apps you pay through your plan instead, and the model you chat with is one factor in how fast you use up your usage limit. So choosing a family means choosing how much capability each reply gets, and what you give up for it. That rules out any answer that treats the families as one model at three prices.
Anthropic's models overview and its guide to choosing a model describe each family by the work it is built for:
The fastest and lowest-cost family. Built for high-volume tasks where a quick reply matters, such as classification (sorting items into set categories), extraction and routing.
Described forQuick, simple tasks repeated many times.
Fast, and capable enough for most everyday work. The docs list data analysis and content creation among its uses, and Anthropic calls it a faster, lower-cost complement to Opus.
Described forEveryday writing, analysis and summaries.
The most capable of the three, at moderate speed and the highest cost. Built for long-running knowledge work, complex reasoning and tasks that need nuanced understanding.
Described forHard, multi-step work where depth matters most.
Each family also has newer and older models. A model's full name is its family plus a version number, and the models overview lists the current model in each family next to older ones that are still available. Version numbers change with every release, so this course names families only, and says "the newest Sonnet model" for the current one. On the exam, tell the families apart by their place in the trade-off, never by a version number.
Interactive · one prompt, three families
Choose the family in the selector and follow the operations team's prompt to the reply it gets.
Haiku: the reply arrives fastest and at the lowest cost, but it is brief. It lists open jobs by type and flags nothing for action.
The newest Sonnet model: still quick, and capable enough to read across the counts, so the reply flags the damp and mould wait as the first thing to act on.
Opus: the most capable family covers the same points in more depth, but more slowly and at a higher cost than a weekly summary needs.
KnowledgeSeeing and switching the model in the apps
In the Claude apps the model is a setting of the conversation, not of your prompt. The model you are chatting with is shown next to the send button on the web and desktop, and at the top of the screen on mobile. You can change it at any point in a chat, and the change applies from Claude's next response. So a switch made for one quick job stays in place for every later message in that chat, until you switch back. Which families you can pick depends on the plan: the Free plan offers Sonnet and Haiku, while paid plans, Team included, add Opus. An administrator may also have set the model every new chat starts on, but you can still change it within a chat. So when a reply surprises you, read the model name before you change the prompt:
- Read the model name next to the send button, or at the top of the screen on mobile.
- Click it to open the model selector; "More models" shows further options.
- Choose the family the task was meant for.
- Send the request again. The new family answers from the next response on.
SkillWorked example: one prompt, two replies
Two officers in Wrenfield's operations team, officer A and officer B, each asked Claude to summarise this week's export of the repairs backlog for Monday's team meeting. The export holds job type, estate, date reported and status, and no tenant names or addresses. They sent the same prompt in their own chats, and the replies came back very different:
# the same prompt in two chats Summarise this week's repairs backlog export for Monday's team meeting. Flag anything that needs action. # officer A's chat, selector: Haiku → open jobs by type, nothing flagged # officer B's chat, selector: the newest Sonnet model → the same counts, plus # "damp and mould: 112 jobs, median wait 41 days, act on these first"
Officer A first thought that officer B had written a better prompt, but the two prompts were word for word the same. The cause sat next to the send button. Earlier that morning officer A had switched that chat to Haiku to sort a long list of contractor job codes quickly, and every later message in that chat went to Haiku too. Neither reply was a fault: Haiku gave the fast, short answer it is built for, while the summary needed a judgement about which jobs matter most, the everyday analysis Sonnet is described for. Officer A switched back to the newest Sonnet model and sent the prompt again, and the team now reads the model name before comparing replies. The team also noted where Opus fits: their hardest work, such as checking the regulator's 38-page complaints guidance against Wrenfield's own process. Choosing a family for each task by cost, speed and quality is 3.3.
More precise than the guide. Anthropic's models overview lists a fourth current family, Fable, above Opus, for demanding reasoning and long-horizon work. In the apps it is on paid plans only. On the Team plan, premium seats include it within part of their weekly limits, and standard seats pay for it with usage credits (checked 2026-10-15). Exam answer: the guide's three families, with Opus as the most capable of them.
- Haiku, Sonnet and Opus trade capability against speed and cost: Opus is the most capable, Haiku the fastest and lowest-cost, and Sonnet the balance between them.
- In the apps the model is a setting of each chat, shown next to the send button, and a switch lasts until you change it back.
- On the exam, compare families by their place in the trade-off; a version number tells you only which model in a family is newer.