Skip to content

Mastering Claude (Anthropic) · Claude foundations · lesson 1 of 20 · 14 min

How Claude works and choosing the right model

The mental model: a very capable reader with no memory of its own

Claude is a family of large language models built by Anthropic. Every answer it gives is produced from two things only: what the model learned during training, and what is in its context window for this conversation. The context window holds your messages, any files you attach, your instructions and preferences, remembered details (if memory is on), and the results of tools such as web search or connectors. If a fact is in neither place, Claude cannot know it, however confident the answer sounds.

That single idea explains most surprises:

  • Out-of-date answers happen because training data has a cut-off. Web search, Research or files you provide close the gap.
  • Generic answers happen because the context is thin. Add audience, goal, examples and constraints.
  • Muddled answers in long chats happen because an enormous context full of old detours dilutes focus. Start a fresh chat with a short summary.

Current Claude models have very large context windows (up to one million tokens on the newest API models, which is several long books), but "can fit" is not the same as "will use perfectly". Long documents still reward targeted questions.

The model families in September 2026

Anthropic ships models in named tiers. The names change with each generation, so learn the pattern, then check the model picker and Anthropic's model overview page for the current line-up.

| Tier | What it is for | Examples (Sept 2026) | |---|---|---| | Haiku | Fastest and cheapest; high-volume classification, extraction, short replies | Claude Haiku 4.5 | | Sonnet | The balanced everyday workhorse; strong writing, coding and analysis at moderate cost | Claude Sonnet 5 | | Opus | Most capable mainstream tier; hard reasoning, long agentic tasks, complex coding | Claude Opus 5.5 (launched 22 September 2026), Claude Opus 5 | | Fable | Anthropic's most capable widely released tier, priced above Opus, for the most demanding long-horizon work | Claude Fable 5.1 |

Anthropic also runs restricted-access research models (for example the Mythos line under its Project Glasswing programme) that most users will never see in the picker. Which models you can choose depends on your plan, and on Team and Enterprise plans an admin can limit which models members can use.

Thinking and effort

Newer Claude models can think before answering: they reason through a problem internally, then write the final response. In the Claude apps this shows up as a thinking or extended thinking option; in the API it is controlled with adaptive thinking and an effort setting (low, medium, high, extra high, max). Higher effort means more careful answers on hard problems at the price of time and usage. For a caption rewrite, thinking is wasted; for a pricing model with twelve constraints, it pays for itself. Module 4 covers this in depth.

How to choose, in practice

Use this decision rule for any task:

  1. Is it simple, repetitive or high-volume? (tagging comments, rewriting subject lines, extracting fields) Start with a fast tier.
  2. Is it everyday knowledge work? (drafting, summarising, analysing a report, most coding) Use the balanced default.
  3. Is it complex, high-stakes or long-running? (strategy with trade-offs, multi-document legal or financial analysis, agentic coding across a codebase) Use the most capable tier, with thinking on.
  4. Still weak? Improve the prompt before upgrading the model. Most poor answers are context problems, not model problems.

Worked example: one agency, three tiers

A 12-person social agency in Dubai runs three recurring jobs:

  • 5,000 comments a week need sentiment tags and a "needs reply" flag. They use a fast tier through the API with a fixed label list and spot-check 50 comments each Monday. Cost stays low and quality is measurable.
  • Weekly client reports combine analytics exports with commentary. The balanced tier in a client Project writes the first draft; an account manager edits.
  • A quarterly pricing review weighs retainer models, freelancer costs and churn risk. The strategist uses the most capable model with extended thinking, supplies all the numbers, and asks for assumptions to be listed so she can challenge them.

The lesson: tier choice is a per-task decision, not a loyalty choice.

Hands-on: run a model comparison you can repeat

Pick one real task you do weekly. Use this prompt in two different models from the picker, in fresh chats, with the same attachment.

<context>
I am a [role] at a [type of business] in [country]. The audience for this output is [who].
</context>

<task>
[Describe the job in one or two sentences.]
</task>

<constraints>
- Use only the attached material and say "not in the material" when something is missing.
- Keep it under [N] words.
- British spelling.
</constraints>

<format>
[Table / bullets / email / headings]
</format>

Score each output from 1 to 5 on accuracy, usefulness and editing time needed, and note how long each took. Keep the table: it becomes your personal model guide, and you can re-run it whenever Anthropic ships a new model.

Pitfalls

  • Assuming the biggest model is always best. It is slower and burns usage limits faster. For simple work it rarely changes the result.
  • Trusting model names from old blog posts. Tiers are renamed and retired; the model picker and Anthropic's docs are the source of truth.
  • Blaming the model for missing context. If the answer lacks your pricing, it is because you did not provide your pricing.

How to measure success

You are choosing well when your default model handles most tasks without rework, you consciously move up a tier for hard problems, and your comparison table shows the time and quality difference in your own work rather than in someone else's benchmark.

Video lecture: How Claude works and choosing the right model

Lecture coming soon · 14 chapters · about 10 minutes. Read the full transcript below.

  1. How Claude works
  2. Why this matters
  3. Two sources of knowledge
  4. Three classic surprises
  5. The model families
  6. Thinking and effort
  7. The four-question rule
  8. One agency, three tiers
  9. Common traps
  10. Simple example
  11. Try this now
  12. Watch me do it, part 1
  13. Watch me do it, part 2
  14. Your next step

Lecture transcript

How Claude works

Here is a question that will save you hours. When Claude gets something wrong, is it the model, or is it you? By the end of this lecture you will be able to answer that in seconds, and you will know exactly which Claude model to reach for, for any task, without guessing or overspending.

Why this matters

Why does this matter? Because two expensive mistakes hide here. The first is overspending, sending simple, repetitive jobs to the most powerful model and burning through your limits or your budget. The second is underpowering, asking a fast model to untangle a complex decision and then trusting a shallow answer. Think of it like hiring. You would not ask your most senior strategist to rename five hundred files, and you would not ask a new intern to restructure your pricing. The skill is matching the job to the right level of expertise, and, just as importantly, diagnosing why an answer was weak before you decide what to change.

Two sources of knowledge

Every answer Claude gives comes from exactly two places. The first is training, everything the model learned before a cut-off date. The second is the context window, which is everything in this conversation right now. Your messages, the files you attach, your preferences, remembered details if memory is switched on, and anything fetched by tools like web search or a connector to your Google Drive. If a fact is in neither place, Claude cannot know it. It may still produce a confident sentence, and that is exactly why we check.

Three classic surprises

That one idea explains the three most common complaints. Out of date answers? Training has a cut-off, so turn on web search or give Claude the source. Generic answers? The context is thin, so add your audience, your goal and an example of good work. Muddled answers in a very long chat? The context is full of old detours. Start a fresh chat and paste a five line summary of what matters. Notice that none of these fixes involve changing the model.

The model families

Now the models. Anthropic names its tiers, and the pattern has been stable even as version numbers change. Haiku is the fast, low cost tier for high volume work like tagging or extraction. Sonnet is the balanced everyday workhorse. Opus is the most capable mainstream tier, built for hard reasoning, complex coding and long agent tasks. And in twenty twenty six Anthropic added Fable, its most capable widely released tier, priced above Opus. As of September twenty twenty six the line up includes Claude Haiku four point five, Sonnet five, Opus five point five and Fable five point one. Those numbers will move. The ladder will not.

Thinking and effort

Newer Claude models can also think before they answer. In the apps you will see an extended thinking option. For developers it is an effort setting, from low up to max. Think of effort like asking a colleague for a quick opinion versus asking them to go away and work it through properly. For a caption rewrite, the quick opinion is fine. For a pricing model with twelve constraints, you want the worked version, and you are happy to wait for it.

The four-question rule

Here is the rule I want you to remember. Is the task simple, repetitive or high volume? Start fast. Is it everyday knowledge work, drafting, summarising, most coding? Use the balanced default. Is it complex, high stakes or long running? Go to the most capable model and switch thinking on. And if the answer is still weak, fix the prompt before you upgrade anything. In practice, most bad answers are context problems wearing a model costume.

One agency, three tiers

Picture a twelve person social agency in Dubai. Every week it tags five thousand comments for sentiment, and that runs on a fast tier through the API with a fixed label list and a human spot check of fifty comments. Weekly client reports are drafted by the balanced model inside a client Project, then edited by an account manager. Once a quarter the strategist reviews pricing, weighing retainers, freelancer costs and churn risk, and she uses the most capable model with thinking on, asking it to list every assumption so she can challenge them. Same company, three different choices, each one deliberate.

Common traps

Three traps to avoid. First, assuming the biggest model is always best. It is slower and it uses up your limits faster, and for a simple rewrite it rarely changes the result. Second, trusting model names from an old blog post or video. Tiers get renamed and retired, so the model picker and Anthropic's own documentation are your source of truth. Third, blaming the model for missing context. If the answer does not mention your pricing, it is almost always because nobody gave Claude your pricing. You will know you are choosing well when your default model handles most tasks without rework, and you move up a tier deliberately, not out of habit.

Simple example

Let's start with a simple example. Take one caption, say, new autumn menu launches Friday, come early for the pumpkin latte, and ask a fast model and the most capable model to make it friendlier. The two answers will be almost identical, and the fast one arrives first. Now take a harder task. Compare three delivery pricing options for a small bakery, with fuel costs, minimum order values and a target margin. Here the gap opens up. The more capable model, with thinking on, lays out a table, spots that one option loses money on small orders, and lists its assumptions. The fast model gives a tidy paragraph that misses the loss. Same person, same afternoon, two different right answers about which model to use.

Try this now

Try this now. Pause the video and pick one task you do every week, a report summary, a client email, a data check. Open two fresh chats. In the first, choose a fast or balanced model. In the second, the most capable model with thinking on. Paste exactly the same prompt and the same attachment into both. Then score each answer from one to five for quality, and write down how many minutes it would take you to make it usable. If the scores are close, the faster model wins for that task. If the capable model saves you ten minutes of fixing, it wins. Either way, you now have evidence, not a hunch, and it took less than fifteen minutes.

Watch me do it, part 1

Let me show you exactly how I run this comparison. I open two fresh chats side by side. In the left one I open the model picker and choose a fast model. In the right one I choose the most capable model and switch extended thinking on. Now I paste the comparison prompt from the lesson into both. Context, I am the operations lead at a bakery chain in Leeds and this goes to our finance director. Task, compare the three delivery options in the attached quote. Constraints, use only the attached material, say not in the material when something is missing, under two hundred words. Format, a table then a one line recommendation. Then I attach exactly the same PDF quote to both chats and press send in each.

Watch me do it, part 2

Now I read both answers the way a finance director would. The fast answer is tidy, but it ignores the surcharge for orders under fifty pounds, so I mark it amber. The capable model builds a table, flags that option two loses money on small orders, and lists its assumptions. I score them. Quality, three for the fast answer, five for the capable one. Minutes to make it usable, about twelve versus three. For this task, the capable model wins, and I write that in my model guide. Then I run a second task, rewriting a product caption, and the scores come out equal, so the fast model wins that one. Two tasks, two evidence based decisions, in about fifteen minutes.

Your next step

So, to recap. Claude knows only its training plus what is in the context. The model families run from fast to most capable, and names change, so trust the picker, not old blog posts. Use the four question rule, and fix prompts before you upgrade. Your next step is in the lesson text: a comparison prompt you run on two models with the same real task. Score both, keep the table, and re-run it whenever a new model arrives. That table becomes your personal, evidence based model guide.

Key takeaways

  • Claude only works with training plus what is in its context: messages, files, instructions, memory and tool results.
  • Tiers run from fast (Haiku) through balanced (Sonnet) to most capable (Opus) and premium frontier (Fable); names change, so check the picker.
  • Choose per task: fast for volume, balanced for everyday work, most capable with thinking for complex, high-stakes problems.
  • If results are weak, improve the prompt first, then compare models on your own tasks.

Try it

Pick one real task you do weekly. Run it on two different Claude model tiers with the same prompt and note differences in quality, speed and usefulness in a short table.