The model lineup and its limits

Model selection is 16.8 percent of the exam and the fastest marks in it are pure recall. Learn the table. It is genuinely one table, and the current models mostly agree with each other, so the memorisation load is smaller than it looks.

The current lineup

Four current models, and they share their limits, which makes this easier than it looks.

claude-fable-5 and claude-mythos-5, at 10 dollars in and 50 dollars out per million tokens. Mythos is invite-only, which is a real access constraint and not a footnote: you cannot design around a model you cannot call.

claude-opus-5, at 5 dollars in and 25 dollars out.

claude-sonnet-5, at 2 dollars in and 10 dollars out through 2026-08-31, then 3 dollars and 15 dollars. That date is a promotional cliff sitting inside your cost model.

All four run 1M context and 128k output.

The outlier and the legacy shelf

claude-haiku-4-5-20251001 is the cheap one at 1 dollar in and 5 dollars out. It differs on two axes and both are testable: 200k context, 64k output, and no adaptive thinking.

Still active, still callable, and worth recognising by name: claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6, claude-sonnet-4-5-20250929, claude-opus-4-5-20251101. Opus 4.x prices at 5 dollars in and 25 dollars out, matching Opus 5. Sonnet 4.6 and 4.5 are 3 dollars in and 15 dollars out.

Active is not the same as advisable. Several of these legacy ids are precisely where the hard 400s live, which is module five.

Context and output are two different limits

A 1M context window is how much you can put in. A 128k output limit is how much can come out of one response. They are unrelated numbers and conflating them is a classic error.

max_tokens is bounded by the output limit, not the context window. Asking a 1M-context model for 500k tokens of output does not work, because output tops out at 128k.

One boundary on that number, worth knowing so the rule stays exactly true rather than roughly true: 128k is the ceiling on the synchronous Messages API. The Batches API raises it to 300k on Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 5 and Sonnet 4.6 behind a beta header. That does not soften the point above, since the request being judged is almost always a plain Messages call, but "128k, full stop" is the kind of half-true that a well-built question will happily charge you for.

No long-context premium

Dated ids versus current ids

claude-sonnet-5 has no date. claude-haiku-4-5-20251001 does. The dated form pins you to an exact snapshot, which is what you want for reproducibility and evaluation, and it is also the form that carries a retirement date.

There is a third form worth naming because it bites hardest: a bare family alias in a config file, the kind of thing that reads as opus rather than as a model id. An alias is a redirect owned by whatever resolves it, so it can silently re-point on any update while your config file still reads exactly as it did yesterday, and the failure is invisible because nothing errors. If you have ever discovered that a service was running a generation behind what its config appeared to say, this is the mechanism.

So: pin dated ids in anything whose behaviour you have measured. Then put the retirement date in a calendar, because nothing else will remind you. claude-opus-4-1-20250805 retires on 2026-08-05, and that is what a retirement looks like arriving.

{
  "model": "claude-sonnet-5",
  "max_tokens": 300000,
  "messages": [
    { "role": "user", "content": "Rewrite this 200-page manual in full." }
  ]
}