Four models in four weeks: September 2026 for Claude developers

A month of releases

In the space of one month, Anthropic shipped four models. Fable 5.1 replaced Fable 5 at the top of the widely available lineup, with Mythos 5.1 — the same model, offered only to Project Glasswing organisations — released alongside it. Opus 5.5 followed as the new default Opus, and Sonnet 5.5 closed out the month as the new Sonnet. Around them, Sonnet 5's introductory price became its permanent one, preserved-thinking enforcement began for new accounts, and paid plans got higher usage limits. The change log has every event with its date. This post is about what the month adds up to for someone maintaining code against the API.

The same breaking changes, three times

Read the three migration guides back to back and the striking thing is how much they share. Fable 5.1, Opus 5.5 and Sonnet 5.5 all reject forced tool_choice. All three bind thinking blocks to the conversation that produced them, so editing earlier history invalidates later reasoning. On all three, text the model writes between tool calls can arrive as thinking blocks rather than text, which silences any interface that renders only text. The two newer ones only accept the computer toolset on the Claude API and Google Cloud. And all three make thinking harder to switch off: always on for Fable 5.1 and Opus 5.5, with a between_tools floor replacing disabled on Sonnet 5.5.

That shared shape is the most useful fact about the month. A harness made compatible with one of these models is most of the way to all three. The work is the same work: an append-only history, tool steering in the prompt instead of forced tool choice, and a client that renders thinking blocks. Do it once, properly, and the next model built on the same mechanics should be much closer to a model-id change and an evaluation run than to another round of firefighting.

What each model is for

Opus 5.5 is the one most teams should move to. Anthropic's migration guidance names it the default Opus target, positions it as the default for most work including complex agentic coding, and prices it below Opus 5 per token. Its effort default is lower than Opus 5's, which changes a bill without changing any code — worth understanding before you switch, and covered in Opus 5.5's medium default.

Sonnet 5.5 is the straightforward upgrade for Sonnet workloads: the same prices and tokenizer as Sonnet 5, a lower minimum cacheable prompt, recalibrated effort levels and a wider set of features. The cost of the move is entirely in the request code, and what changed from Sonnet 5 lists it.

Fable 5.1 is for when you have chosen it deliberately. Anthropic reserves it for the toughest long-running agentic and research problems, and for work Opus 5.5 can't handle even with more effort — it is explicitly not the default upgrade path. It keeps Fable 5's per-token prices but reads its cache at a far deeper discount, which makes long cached sessions noticeably cheaper than before.

Mythos 5.1 only matters if your organisation is approved for Project Glasswing. For everyone else, Fable 5.1 is the top of the lineup.

Haiku didn't get a new model this month. Anthropic has said a Haiku 5.5 will join the family, and until it ships, Haiku 4.5 remains the small, fast option.

The cost story is in the details

The per-token headlines were modest: Opus 5.5 cheaper than Opus 5, everything else holding its price. The interesting changes sit one column over. Cache reads stopped costing the same fraction of input on every model, so a comparison without the cache-read rate now misleads. Effort defaults and level calibrations moved, so the same setting no longer buys the same amount of thinking. And thinking got harder to switch off, which lands on routes that used to run without it. None of these appears on a rate card as a price change, and all of them move invoices.

The month also produced one price that didn't change when it was supposed to — and this site got it wrong. Sonnet 5 kept its launch price is the correction.

Before you migrate anything

A short list that covers most of the month's breakage:

• Check stop_reason for "refusal" before reading content, and opt into fallbacks from the first deploy — handling refusals has the setup.
• Make the conversation history append-only, and run the preserved-thinking check with the binding field set, even if your own account isn't enforced.
• Replace forced tool_choice with prompt steering, strict tool use and a check that the call happened.
• Render thinking blocks, with the "updates" display where it's offered, so long turns don't go silent.
• Set effort explicitly on every route and re-run your effort sweep; level names don't carry across models.
• Size max_tokens for thinking as well as the answer.
• Watch cache_read_input_tokens after the switch. On the models with the deepest read discounts, a cache that silently stopped hitting is one of the most expensive bugs you can ship.