Claude Sonnet 4.5 Guide: Retirement Date and Where to Migrate

If you still have claude-sonnet-4-5 in production, the most useful fact about it is a date. Anthropic deprecated Sonnet 4.5 on 30 September 2026 and lists its retirement on the Claude API as 30 November 2026. After that, requests to it on Anthropic-operated platforms fail.

That leaves roughly six weeks from the time of writing. This guide covers what Sonnet 4.5 was, which of its behaviours your code may quietly depend on, and how to choose between the two realistic replacements.

The retirement, precisely

Anthropic's deprecation page lists the dated model claude-sonnet-4-5-20250929 as deprecated on 30 September 2026, with retirement on 30 November 2026 and claude-sonnet-5-5 as the recommended replacement. Three details matter:

• The dates apply to Anthropic-operated platforms: the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own retirement schedules, so a model can live longer there.
• Third-party routers depend on the upstream providers they route to. Do not assume a router will keep serving the model after Anthropic stops.
• Deprecated models can be less reliable than active ones in the meantime, by Anthropic's own warning.

The Console's usage export breaks traffic down by API key and model, which is the quickest way to find stragglers. Model deprecation and migration covers running a sunset without a scramble.

What Sonnet 4.5 was

Anthropic released Sonnet 4.5 on 29 September 2025 at $3 / $15 per million tokens and called it the best coding model in the world at the time. Its launch numbers were 77.2% on SWE-bench Verified (averaged over ten trials, rising to 82.0% with a high-compute setup) and 61.4% on OSWorld, up from 42.2% for Sonnet 4. Anthropic also said it had observed the model maintaining focus for more than 30 hours on complex multi-step tasks.

The launch shipped with infrastructure that shaped later work: the Claude Agent SDK, context editing and a memory tool in the API, and checkpoints in Claude Code. It was released under Anthropic's ASL-3 protections.

On limits, Anthropic's documentation lists a 200K-token context window. OpenRouter's catalogue shows 1M, reflecting a long-context configuration offered at launch; plan around 200K, which is what Anthropic now documents. OpenRouter lists maximum output at 64K tokens.

Behaviours your code may depend on

Sonnet 4.5 predates several API changes, so code written for it often uses things the replacements reject or treat differently. Audit for these before you swap the model string:

• Assistant-turn prefill. Prefilling the start of the reply, often used to force a JSON shape, works on Sonnet 4.5 and returns a 400 on Sonnet 4.6 and later.
• Thinking budgets. Sonnet 4.5 uses thinking: {type: "enabled", budget_tokens: N}. Sonnet 4.6 still accepts it but deprecates it; Sonnet 5.5 rejects it.
• No effort parameter. Sending effort errors on Sonnet 4.5. Sonnet 4.6 defaults it to high, which can raise latency and token use if you do not set it.
• Sampling parameters. temperature and friends work on Sonnet 4.5 and 4.6, and return a 400 on Sonnet 5.5.
• Interleaved thinking beta header. Sonnet 4.5 needs it; models with adaptive thinking do it automatically, so remove the header.

Sonnet 4.6 or Sonnet 5.5?

Both are reasonable landing spots, and they trade effort now against effort later.

Sonnet 4.6 is the smaller step. Same $3 / $15 price, same tokenizer, so token counts and cost models carry over. It keeps sampling parameters, forced tool use and disabled thinking. The work is mostly removing prefill and setting effort. Its context grows to 1M tokens. Anthropic lists its retirement as not sooner than 17 February 2027, so you buy time rather than a permanent home. The Sonnet 4.6 guide covers it in detail.

Sonnet 5.5 is Anthropic's recommended replacement. It lists at $2 / $10, but uses the newer tokenizer that produces roughly 30% more tokens for the same text, so check real token counts before assuming the saving. It also rejects sampling parameters, thinking budgets, forced tool use and disabled thinking. That is more rework now, but no second migration in a few months. The Sonnet 5.5 guide covers its specifics.

A practical rule: if your integration is simple or you have the time, go straight to Sonnet 5.5. If it leans on prefill plus temperature plus forced tool calls and the deadline is close, hop to Sonnet 4.6 now and schedule the larger rewrite.

A migration plan that fits six weeks

1. Inventory. Export usage by key and model; grep the codebase for claude-sonnet-4-5 and the dated ID.
2. Audit requests for prefill, budget_tokens, sampling parameters and beta headers.
3. Pick a target per route. Different routes can land on different models.
4. Re-baseline. Run twenty real tasks per route on the target and compare acceptance rate, tokens and latency against Sonnet 4.5. How to benchmark LLMs on your own work has a harness.
5. Pin the new ID and ship well before 30 November, leaving room to roll back.

Variants

The catalogue lists the standard endpoint and a :batch variant, which is the same model on Anthropic's asynchronous Batch API at half price, $1.50 / $7.50 per million tokens. Both retire with the model. If you pinned the dated ID claude-sonnet-4-5-20250929 to keep behaviour fixed, that was the right instinct, and the same habit applies to whichever replacement you choose.