Almost everyone picks a model the same way. They read that one is the most capable, use it for everything, and then get surprised by a bill. Or they pick the cheapest, get a run of poor answers, and conclude the whole category is overrated.
Both come from treating this as a ranking question. It is a cost question with a judgement attached, and once you see the actual numbers the decision gets much easier to make.
What the three tiers cost, and the ratio that matters
Prices below are per million tokens, from Anthropic's published pricing page and correct on 29 August 2026. Check the page rather than this post before making a budget: these move.
- Claude Sonnet 5: 2 dollars in, 10 dollars out.
- Claude Opus 5: 5 dollars in, 25 dollars out.
- Claude Fable 5: 10 dollars in, 50 dollars out.
The absolute numbers matter less than the shape. Opus costs two and a half times Sonnet, Fable costs twice Opus, and Fable costs five times Sonnet, on both input and output, all the way up.
Five times is the number to hold on to. It is small enough that it should never stop you using the strongest model on something that genuinely matters, and large enough that it will hurt if you use it for everything by default.
One recent change worth knowing if you priced this earlier in the year: Sonnet 5's two-and-ten pricing was announced as introductory and scheduled to rise on 1 September 2026. Anthropic has since confirmed that increase will not happen and the introductory rate is now the standard price.
What actually changes as you move up the tiers
The unhelpful description is that the expensive ones are smarter. The useful one is that the failure mode moves.
On a straightforward task with a clear answer, the tiers converge. Summarise this, classify that, extract these fields, rewrite this paragraph. A cheaper model does those about as well, and paying five times more buys you very little.
The gap opens on work where the difficulty is not in the answer but in getting there. Long chains of reasoning where an early wrong turn is unrecoverable. Problems that need holding a lot of context at once. Agentic work that runs for many steps, where a small misjudgement compounds instead of showing up immediately.
That is the real distinction. The stronger model is not producing better sentences. It is going wrong less often on the kind of problem where going wrong is expensive.
A cost trap that catches people comparing prices
Per-token prices are only comparable when the tokens are the same size, and across model generations they are not.
Anthropic notes that Claude 4.7 and later models use a newer tokenizer which produces roughly thirty percent more tokens for the same text than the one used by Sonnet 4.6 and earlier. The exact increase depends on the content.
So a per-token price cut between generations can be smaller in practice than it looks, because the same document now costs more tokens. Comparing Sonnet 5, Opus 5 and Fable 5 against each other is fair, since they share a tokenizer. Comparing any of them against an older model on price alone is not.
This is the sort of thing that turns a projected saving into an unexpected invoice, and it is invisible unless somebody tells you.
How to choose, in practice
Anthropic's own guidance is a reasonable default: the smallest model for simple tasks, the middle tier for most production work, and the top of the range for the most complex reasoning. In practice one question does most of the work.
Ask what it costs when this answer is wrong and nobody notices. If a bad output gets caught immediately by a person or a test, use the cheap model, because you are paying for a mistake that gets found. If a bad output flows into a decision, a customer, or a long chain of further steps, buy the better model, because the five times difference is trivial next to the cost of the failure.
- Start on the middle tier for anything real, and move in whichever direction the results push you.
- Move down when you cannot tell the difference on your actual work. Test that with your own tasks rather than assuming it.
- Move up for long agentic runs, hard reasoning, and anything where an early mistake is not recoverable.
- Route rather than standardise. Cheap model for the bulk classification, strong model for the twenty decisions a week that matter.
Two cheaper levers than downgrading
People reach for a smaller model when the bill arrives, and it is usually the third-best option. Two others cut cost without giving up capability.
Prompt caching charges a cache read at a tenth of the standard input price, so anything you resend on every request, a long system prompt, a document, a conversation history, is repeatedly billed at full price until you cache it.
The Batch API takes fifty percent off both input and output for work that does not need an answer immediately. Overnight jobs, bulk classification and backfills are all a straightforward halving.
Cache and batch first. A tenth on repeated input beats moving down a tier, and it costs you nothing in quality.
The part that is still yours
None of this tells you which model is right for your work, because that depends on what your work punishes. The tiers give you a dial between cost and the chance of a confident mistake, and where to set it is a judgement about consequences rather than a fact about the models.
Which is the same thing this blog keeps arriving at. The tooling made producing cheap. Deciding what a wrong answer costs you is the part that did not move.