Claude Sonnet vs Opus vs Fable: which to pick

7 min read · every claim from Anthropic’s own pages, linked below

The short version: start on Sonnet. Anthropic’s own guide calls it the daily driver. Opus is the slow thinker, for the problem you keep going back and forth on. Fable is built for jobs that run for hours, where you hand over a whole project and walk away. Every step up uses your limit faster, and Fable uses the most. On the Pro plan, Fable is not covered by your plan at all: it bills as pay-as-you-go credits from the first message. On Max it is included, up to half your weekly limit.

First, what the menu actually is

Open Claude and look next to the send button. There is a model name, and if you click it, a short menu: Sonnet, Opus, Fable, and under “More models” a few others including Haiku. That menu is the model picker.

Think of it as the same assistant with different size brains. They all read your files, search the web and write in the same voice. What changes is how long the model thinks, how much it can hold together at once, how fast it answers, and how much of your usage it burns doing it.

The instinct is to tap the biggest one and leave it there. That is the wrong move, and the reason is not complicated: the bigger the model, the faster your limit goes, and for most of what you ask in a day the answer is no better.

Which model for which job

This is how Anthropic itself describes the four, in its Claude Academy guide and its model documentation, with the kind of task each one is for.

Sonnet: the one you should be on most of the day

Say you need an email rewritten so it sounds less annoyed. Or somebody sent you a spreadsheet and you want to know what it says. Or you are drafting a post, fixing a function, summarising a report. That is Sonnet. Anthropic’s documentation describes it as “the best combination of speed and intelligence”, and the Academy guide calls it your versatile default. It answers quickly and it is moderate on your limit.

Opus: the slow thinker

Opus is for the problem that does not fit in one question. The decision you have been chewing on all week. The contract with three clauses that seem to contradict each other. The bug nobody can find. The session where you go back and forth for an hour and need it to keep the whole thread straight. Anthropic’s guide puts it as “problems that genuinely need deep thinking over time”. It is heavier on your limit, and it needs a Pro plan or above.

Fable: the one you hand a project to

Fable is the newest and the most capable model Anthropic offers to the public. But the word that matters in its description is not “smartest”, it is hours. Anthropic says Fable 5.1 is “best for ambitious, long-running, asynchronous work” and “built for jobs that take hours and span many applications”. Its documentation gives the examples: agent sessions that run for hours, multistep deep research, analysis carried through to a finished document, spreadsheet or deck.

In plain terms: you hand over the whole project, walk away, and come back to something finished. If what you have is a question, Fable is the wrong tool. It is slower, and you are paying its rate for an answer Sonnet would have given you.

Why not just use Fable for everything

Because every step up eats your limit faster. Your plan gives you a usage allowance, and the models draw it down at different rates. Anthropic’s Academy guide lays the order out plainly: Haiku is the lightest, Sonnet is moderate, Opus is heavy, and Fable uses “the most of your rate limit”.

Anthropic does not publish the exact ratios for the app, so there is no honest number to give you there. The closest public proxy is what developers pay per million tokens (the chunks of text a model reads and writes), which is a fair picture of how much heavier each step is:

Those are API prices, read on 18 September 2026, not what your subscription charges. They are here because they show the shape: the same question costs roughly five times as much to answer on Fable as on Sonnet. Spend that on an email rewrite and you have used your week on nothing.

What your plan really gets you

This is the part that catches people, because Fable shows up in the picker on every paid plan and nothing in the menu tells you it bills differently.

The wording on Anthropic’s help page is direct. On Pro and standard seats the Fable models “aren’t included in your plan’s usage limits”. On Max and premium seats “you can use up to 50% of your weekly usage limits on Fable models at no extra cost”.

If you remember Fable being free on Pro, you are not imagining it. There was a launch promotion for Fable 5, and the same page says it ended on 19 July 2026.

What Fable hands off

One more thing the picker does not tell you. Fable is capable enough in a few areas that Anthropic put safeguards around them, and when a request trips one, Fable does not answer it. The request falls back to Opus instead.

The areas listed on Anthropic’s help page:

1. Offensive cybersecurity. Building exploits, malware or attack tooling. These fall back to Opus 4.8.

2. Dual-use biology. Virology, toxicology, drug design and molecular design. Biology, chemistry and life sciences requests fall back to Opus 5.

3. Distillation attacks. Attempts to extract the model’s thinking or capabilities.

4. Frontier AI development. Distributed training, ML accelerators and specialised kernel development.

It does tell you. You get a notice that the model switched, and the response is labelled with the model that answered. The detail that is easy to miss: the picker then stays on Opus for the rest of the conversation. If you work in security or life sciences and wondered why a long Fable session quietly stopped feeling like Fable, this is why. You can switch back from the picker at any time, edit and retry the message, turn automatic switching off in Settings, under Capabilities, or send feedback if the block was wrong.

The rule, in three steps

If your limit is the real problem, the model is only one of the levers. The other habits are in six habits that cut your Claude token usage. And if you are still working out what Claude keeps between conversations, that is what Claude remembers about you, and how to wipe it.