The Claude Model Family

Think of the family as one assistant available in three settings, each tuned differently. Haiku is the fastest and lightest - and the cheapest to run. It answers almost instantly and is perfect for quick, simple, high-volume jobs: tidying a list, reformatting notes, a fast reword, a first rough pass where you do not need deep thought. Ask it a genuinely hard, many-step question and it may come up short - not because it is broken, but because that is not what it is for.

Sonnet is the balanced everyday workhorse, and for most architects and designers it is the sensible default. It is markedly more capable than Haiku at reasoning, drafting and following complex instructions, while staying fast and economical enough for constant use. The great majority of the work in this course - drafting a spec, structuring a brief, comparing options, writing a report, most research - sits comfortably here. If you are unsure which tier to open, open Sonnet.

Opus is the most capable - the deepest reasoner, best on complex, long, high-stakes or intricate tasks: a knotty multi-part analysis, a long document held together as a whole, a subtle piece of code, a problem where the quality of thinking really matters. That depth comes at a price. Opus is slower to respond and costlier to run, and on higher plans heavier use can bump against limits sooner. Reaching for Opus on a task Haiku could do is like sending a principal to photocopy - it works, but it wastes the resource.

The figure lays the three side by side on the axes that matter: speed, depth and cost. Read across and the pattern is simple and always holds, whatever the models are called: as you move Haiku to Sonnet to Opus, depth of reasoning rises, and so do latency and cost. There is no single 'best' model - only the best fit for the task and the stakes in front of you, which is the whole skill this lesson builds.

Why can't one model just be fast, cheap and deep all at once? Because depth of reasoning takes work - more computation per answer - and computation is what turns into time and cost. A larger, more capable model considering a problem more thoroughly simply does more under the hood than a light one firing back a quick reply. So the trade-off is not a marketing gimmick; it is real, and understanding it stops you both from overpaying for trivial tasks and from starving a hard problem of the capability it needs.

Newer models add another dial worth knowing: an extended-thinking or reasoning mode. Turned on, Claude spends longer working a problem through before answering - in effect thinking step by step at greater length - which noticeably improves results on genuinely hard, multi-step tasks: intricate logic, careful analysis, tricky code, problems with several interacting constraints. It trades still more time (and cost) for still more rigour. For a quick reword it is pointless overkill; for a subtle trade-off study or a knotty script it can be the difference between a shallow answer and a sound one. Treat it as a tool you reach for deliberately when depth matters, not a switch you leave on by default.

Cost shows up differently depending on how you use Claude, and it is worth being honest about because this course never sells you anything. On the consumer plans you pay a flat subscription and heavier use of the deepest tier tends to hit usage limits sooner rather than costing per message; through the API, where software calls Claude, you pay by the amount of text processed, so tier choice maps directly to rupees. Either way the instinct is the same: use the lightest tier that does the job well. Reserve Opus and extended thinking for the tasks that genuinely reward them, run everyday work on Sonnet, and let Haiku mop up the quick and simple. Matching capability to the task is not stinginess - it is good practice, faster answers, and often a clearer result, since a heavyweight model can over-think a task that wanted a light touch.

So how do you actually choose, task by task? The decision in the figure is the one to internalise, and it runs on two questions: how hard is the thinking, and how much is at stake? Start with the middle. If it is everyday work - drafting, research, most specs, ordinary correspondence - reach for Sonnet. It is the default for a reason, and you will spend most of your time here.

Step down to Haiku when the task is quick, simple and repetitive - reformatting a list, a fast tidy of notes, extracting fields from a block of text, a first rough pass you will heavily rework anyway. You gain speed and, on the API, real cost savings, and you lose nothing, because the task never needed deep reasoning. Step up to Opus - and consider turning on extended thinking - when the task is complex, long, high-stakes or intricate: a weighted comparison of several schemes with interacting factors, a long rationale that has to hold together as one argument, a subtle script, an analysis where a shallow answer would mislead you. Here the extra time and cost buy quality that matters.

Two practical notes make this easier in daily use. First, when in doubt, start on Sonnet, then adjust - if the answer feels shallow or the model is visibly struggling with the complexity, move up to Opus; if you find yourself using a heavyweight for something trivial, drop to Haiku. You do not have to get it right first time. Second, remember the previous lesson: tier is about depth of reasoning, not about truth. A more capable model hallucinates somewhat less and reasons better, but Opus is still a plausibility engine, not an oracle - it does not lift your obligation to verify high-stakes facts at the source. Choosing the right tier makes Claude a better assistant; it never makes it the architect of record. That job does not have a model.

The reason this lesson talks in tiers rather than version numbers is simple: the numbers churn and the tiers endure. Claude has moved through earlier 3.x and 4.x generations to the Claude 5 generation as of 2026, and it will keep moving - each generation lifting all three tiers. If you tie your habits to 'use model X.Y', your knowledge expires with the next release. If instead you think 'this is a light task, a default task, or a deep task', you choose well no matter what the current models are named. When a new generation lands, you do not relearn anything; the fast one is still the fast one. Pick the tier that fits the task, check the current names when you need to, and let the version race pass you by.

It is also worth knowing the three places Claude runs, because they are not interchangeable. Claude.ai is the chat interface - browser, desktop and mobile app - where you type, paste and attach files, and where an architect or designer does almost all of their work; this whole course lives here. The Anthropic API lets other software call Claude programmatically and bill by usage - it is how tools inside Studio Matrx use Claude behind the scenes, and it matters mainly in Module 8. Claude Code is a version for developers that works in a terminal to help write software; you will only meet it lightly, later. For everything in the modules ahead, Claude.ai is your home, and on it you simply pick the tier - Haiku, Sonnet or Opus - that suits the task in front of you.

Hold on to the shape of it. One family, three tiers, one trade-off: speed and cost against depth. Sonnet for most days, Haiku for the quick and simple, Opus and extended thinking for the genuinely hard - and none of them a substitute for your judgement. Get that, and you are choosing like a professional, this year and every year after, whatever Anthropic decides to call the models next.