Claude Sonnet 4.5 Deprecation: Migrate to Sonnet 5.5

One model ID: claude-sonnet-4-5-20250929. Anthropic emailed affected accounts on 30 September and the model deprecations page lists 30 November 2026 as the retirement date. That's 61 days of notice, one day above the 60-day minimum Anthropic commits to.

The date applies to Anthropic-operated surfaces: the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud run their own retirement schedules, so if you call Sonnet 4.5 through Bedrock you may have longer — or not. Check their model tables instead of assuming.

The part people skip is finding every caller. Sonnet 4.5 was the default Sonnet for most of late 2025, which means it's hard-coded in places you forgot about: an n8n Anthropic node from last year, a .envon a VPS, a Make scenario a client built. The Console has an audit for this — Usage → Export gives you a CSV broken down by API key and model. Do that first, then grep. A quick grep -rn "sonnet-4-5" across my own automation scripts this morning turned up a video pipeline still hard-coding claude-sonnet-4-5-20250929— a script I'd have sworn was on Sonnet 5.

I took a small lead-classification request and a 20,000-character extraction prompt and sent variations of each to both models on 1 October 2026. Everything below returned HTTP 400 on claude-sonnet-5-5. The error strings are copied from the responses.

The prefill one will hit the most people.Sonnet 4.5 was the last Sonnet that let you start Claude's reply with {to force raw JSON. It's a habit from 2024 and it's everywhere. The same test showed why people did it: asked to “output only JSON”, Sonnet 4.5 still wrapped its answer in a ```json fence. Sonnet 5.5 returned bare JSON without being forced. The right replacement is structured outputs (output_config.format) — I covered the trade-offs in structured outputs vs tool use.

Two notes. temperature: 1is accepted because it's the default value; any other number fails, so delete the parameter rather than tuning it. And forced tool use isn't coming back: use tool_choice: auto, mark the tool strict: true, and say in the prompt when to call it. My classifier called the tool in all three runs with just “Use the classify_lead tool” appended.

Per token, yes. Per request, it depends entirely on one setting you probably never sent. Two things move at once.

1. The tokenizer counts more tokens

Sonnet 5.5 uses the newer tokenizer that arrived with Claude 4.7. I ran the token counting endpoint on the same 20,000-character system prompt (a TSX page, so code-heavy): 6,328 tokens on Sonnet 4.5, 7,823 on Sonnet 5.5 — 23.6% more. Anthropic says “about 30%” and that it varies by content. Even so, the rate cut wins on input: 7,823 × $2 is about 18% cheaper than 6,328 × $3.

2. It thinks by default

This is the one with no error. On Sonnet 4.5, a request with no thinking field ran without thinking. On Sonnet 5.5 the same request runs adaptive thinking at high effort. Here's the extraction task, one run each, same prompt:

So a lazy swap of the model ID made this call 29% more expensive and 2.7× slower, on a model that's supposed to be cheaper and faster. One explicit output_config.effort: "medium"flipped it to 14% cheaper and faster than before. It's a single run on a small task, so treat the percentages as direction, not gospel — but the direction is the whole lesson. If you're not sure which level fits which route, I wrote up how I pick an effort level.

Also raise max_tokens. It now covers thinking plus text, and thinking is billed as output. A route that used to fit in 300 tokens can get truncated mid-thought.

• content[0].text breaks. With thinking on, the response starts with a thinking block whose text is empty by default. Read blocks by type. Every quick script I've ever written does content[0].text.
• Narration between tool calls moves. Notes longer than a sentence or two now arrive as progress-update thinking blocks, empty at the default display. A UI that streamed them goes quiet. between_tools brings the text back; with adaptive thinking set display to "summarized".
• Thinking blocks are tied to the conversation.Edit an earlier turn, the system prompt or the tools and replay the block, and newer accounts get a 400. Keep history append-only — which also protects your cache, as I found in cache miss diagnostics.
• Images cost more.Sonnet 5.5 reads up to 2576 px on the long edge; a 2000×1500 image costs about 2.5× the tokens it did on 4.5. Downscale first if you don't need the detail.
• Caching gets easier. The minimum cacheable prompt drops from 1,024 to 512 tokens, so short system prompts that never cached now do.
• Beta headers to delete: interleaved-thinking-2025-05-14, any context-window header, and fine-grained-tool-streaming-2025-05-14 (replace with eager_input_streaming: true per tool). Move output_format to output_config.format.

It's a legitimate stopgap. claude-sonnet-4-6is active until at least 17 February 2027, costs the same $3/$15 as 4.5, keeps the old tokenizer, doesn't think unless you ask, and still accepts temperature and (deprecated) budget_tokens. It does reject prefill. If you have fifty client workflows and two weeks, that's the smallest diff.

My take: don't, unless you're truly out of time.You'd pay 50% more per token for an older model that already has a retirement floor on the calendar, then do this whole migration again next year. Most of the work — prefill, forced tools, parsing by block type — you'd do anyway. I wrote about the same dilemma one tier up in Opus 5.5 vs Opus 5, and the answer was the same: fix it once, properly.

1. Find callers. Console Usage export, then grep -rn "sonnet-4-5" across repos, n8n exports and .env files.
2. Swap the ID to claude-sonnet-5-5 (no date suffix; anthropic.claude-sonnet-5-5 on Bedrock).
3. Kill the 400s. Remove prefills, temperature/top_p/top_k, budget_tokens; replace forced tool_choice with auto + strict: true.
4. Set effort on every route. medium for well-scoped agent steps, low or between_tools for classification and chat, high only where quality measurably needs it.
5. Fix parsing. Read by block type, pass thinking blocks back untouched, keep history append-only, raise max_tokens.
6. Re-baseline.Count tokens on real prompts, compare cost and latency per task, not per million. If it's not latency-sensitive, push it through the Batch API for another 50% off.

If you use Claude Code, /claude-api migrate this project to claude-sonnet-5-5does most of steps 2–4 and hands you a checklist. Review its diff like you'd review a junior's; it doesn't know which routes are latency-sensitive. Anthropic's Sonnet 5.5 migration guide has the full before/after per starting model.

One more thing on timing: don't wait for 29 November. Deprecated models are, in Anthropic's own words, likely to be less reliable than active ones. I'd have production off 4.5 by mid-November and leave the last two weeks for the caller you didn't find.