Claude Haiku 4.5 retires no sooner than Oct 15: Sume's Haiku row

Anthropic's models overview lists Claude Haiku 4.5 with a retirement of "not sooner than October 15, 2026", and Sume's Agents catalog carries a Haiku 4.5 row with no successor configured. If an agent of yours is pinned to Haiku 4.5, plan the move now, because nothing in Sume's registry will silently remap it for you.

"Not sooner than" is Anthropic's wording for a commitment on Anthropic-operated platforms, not a shutdown date. The same page says Amazon Bedrock and Google Cloud set their own dates. Treat October 15 as the earliest day the model can go, and check the deprecations page the overview links to before you rely on a later one.

What does Anthropic's page say about Haiku 4.5?

The overview table compares four current models. These are the Haiku 4.5 cells next to Sonnet 5.5, the Anthropic row Sume added most recently:

• Haiku 4.5 is the only current Claude on that page with a 200K window. Sonnet 5.5 and the other current models list 1M.
• Anthropic's table says effort is not supported on Haiku 4.5, and that its thinking mode is the older extended thinking.

What does Sume do with a Haiku 4.5 pick?

In Sume's model registry the Haiku 4.5 row is a normal catalog row: label "Haiku 4.5", described as Anthropic's fastest efficient model, running on the Anthropic route through the same OpenRouter-style catalog as Sonnet 5.5. It is listed only where that catalog gate is open, which is an environment setting. If your workspace picker does not show it, an API request naming it is refused rather than moved to another model.

The important part is what is missing. Sonnet 5 has a successor entry that sends it to Sonnet 5.5, and Opus 5 sends to Opus 5.5. Haiku 4.5 has no such entry. When Anthropic retires it, nothing in the registry will redirect it, so expect an error from the provider until Sume ships a change. Do not assume a quiet fallback.

How do I move a Haiku-pinned agent before the date?

Pick the replacement on purpose and log which one ran. For Formats, the receipt's model field is the catalog id the orchestrator used, per Runs and results. Run the same instruction and input once on each candidate with a low generation_spend_cap_usd and compare the receipts side by side.

Candidates that Sume lists and Anthropic keeps current: Sonnet 5.5 for the same family at double the list price, or a non-Claude row such as DeepSeek V4.1 Flash if your reason for Haiku was low cost. Whichever you choose, the media models do not change, because the orchestrator only plans and calls tools. The Sonnet versus Opus post covers the tool-calling side of that choice.

What can go wrong?

An id outside the catalog is 400 invalid_request on a Format run, so a typo cannot slip onto a different model. Agent Completions takes no model choice at all, only model: "sume-agent", so there is nothing to migrate there.

The risk is a long context: a thread that fit Haiku's 200K window never needed trimming, and moving to a 1M model changes cost, not correctness. Read the receipt's usage after the first run instead of estimating it.