Which Claude model should you use?

Anthropic's current line-up is four models. Several older ones — Opus 4.7, Opus 4.6, Sonnet 4.6 — remain available as legacy; these four are the ones Anthropic lists as current. From its own documentation, on 17 July 2026:

Sonnet 5's introductory price runs until 31 August 2026. From 1 September it rises to $3 / $15 — a 50% increase, on a dated schedule Anthropic publishes today. See our API pricing comparison.

The thing that is not on the pricing page

Anthropic's newer models use a new tokenizer, and its docs state it produces approximately 30% more tokens for the same text. This applies to Fable 5, Mythos 5, Opus 4.7 and later, and Sonnet 5. Sonnet 4.6 and earlier use the previous tokenizer.

Follow that through, because it has two consequences nobody markets:

Your context window shrank. A 1M-token window is a container measured in tokens, not words. Anthropic's own docs put Fable 5's 1M window at roughly 555,000 words, and the older tokenizer's 1M window at roughly 750,000. Same headline number. About 26% less text fits.

Your bill went up more than the price list says. Pricing is per token. If the same document is now 30% more tokens, it costs 30% more to send — at an unchanged per-token rate. Nothing on the pricing page moved. Your invoice did.

This is not a trick; a better tokenizer can improve quality, and Anthropic documents the change plainly. But if you are migrating from Sonnet 4.6 to Sonnet 5 and budgeting on the per-token rate alone, you will under-forecast. Measure with the actual tokenizer, not the old one.

The two cutoff dates

Anthropic publishes two dates per model, and they are not the same thing:

• Training data cutoff — the broader range of data used.
• Reliable knowledge cutoff — the date through which knowledge is most extensive and reliable.

For Haiku 4.5 these are July 2025 and February 2025 respectively — five months apart. The three larger models are January 2026 for both.

The docs explicitly distinguish these, and the second one is the number that matters. If you need a model that reliably knows about something, "reliable knowledge cutoff" is the honest field. Coverage that quotes only the training cutoff is quoting the flattering number.

Choosing

Anthropic's own guidance is not quite consistent across its pages: the model overview says start with Opus 4.8, and use Fable 5 when you need the highest available capability. The pricing page still gives the older three-tier heuristic — Haiku for simple tasks, Sonnet for most production work, Opus for the hardest reasoning — and does not mention Fable at all.

Our reading:

The last row surprises people: the cheapest model is the one that lists extended thinking. The three larger models list adaptive thinking instead — on Fable 5 it is always on.

The two you probably cannot have

Mythos 5 and Mythos Preview are offered separately for defensive cybersecurity workflows under Anthropic's Project Glasswing. Limited availability, invitation only, no self-serve sign-up. Mythos 5 is documented as sharing Fable 5's pricing at $10 / $50.

If you see Mythos in a "best models ranked" listicle, the author has not tried it.

Two practical traps

Model IDs are pinned snapshots. From the 4.6 generation onward, even the dateless IDs like claude-sonnet-5 are pinned snapshots rather than evergreen pointers to "latest". Do not assume an ID without a date auto-upgrades.

Opus 4.1 retires on 5 August 2026. If it is in your stack, that is a hard date, not a suggestion. Anthropic points migrations at Opus 4.8.