Claude Sonnet 4.5 vs Sonnet 5.5: What Changes If You Switch

Claude Sonnet 4.5 retires on the Claude API on November 30, 2026, and Anthropic's recommended replacement is Claude Sonnet 5.5. The short version: Sonnet 5.5 is cheaper by list price ($2 / $10 per million input/output tokens versus $3 / $15), has a context window five times larger (1 million tokens versus 200,000), outputs twice as much per response (128K tokens versus 64K), and runs noticeably faster — but it also rejects a handful of things Sonnet 4.5 code relies on, including assistant message prefill and forced tool choice. If you're migrating because of the retirement notice, here is every number and every breaking change, sourced directly from Anthropic's own documentation.

What actually happens on November 30, 2026

Anthropic notified developers on September 30, 2026 that Claude Sonnet 4.5 had moved from "active" to "deprecated" status, with a tentative retirement date of November 30, 2026, as recorded on Anthropic's model deprecations page. In Anthropic's own terminology, deprecated means the model still works but Anthropic is no longer recommending it; retired means requests to it start failing outright. We covered the retirement notice itself in detail when it was announced, including why the date carries Anthropic's "tentative" qualifier rather than being locked in.

The November 30 date applies specifically to the platforms Anthropic operates directly: the Claude API, Claude Platform on AWS, and Microsoft Foundry. Amazon Bedrock and the Google-managed version of Claude Models on Google Cloud set their own retirement schedules for the same model, which can run longer than Anthropic's own timeline. If you're calling Sonnet 4.5 through Bedrock or Google-managed Vertex-style deployment, check that platform's own model lifecycle page rather than assuming the November 30 date applies to you.

Pricing: what changes on your bill

Sonnet 5.5 is priced lower than Sonnet 4.5 across every line item on Anthropic's pricing page. Input tokens drop from $3 to $2 per million (a 33% cut), output tokens drop from $15 to $10 per million (also 33%), and prompt-cache pricing drops too: a five-minute cache write falls from $3.75 to $2.50 per million tokens, and cached reads fall from $0.30 to $0.10 per million tokens.

There's a wrinkle worth knowing before you celebrate the savings. Anthropic's Sonnet 5.5 migration guide notes that Sonnet 5.5 shares Sonnet 5's tokenizer, and that "against Claude Sonnet 4.6, Claude Sonnet 4.5, and Claude Haiku 4.5, the same text produces about 30 percent more tokens, depending on the content." So the per-token price is a third lower, but the same prompt and response can consume around 30% more tokens than it did on Sonnet 4.5. Depending on your workload, the net cost change lands somewhere between "noticeably cheaper" and "close to a wash" — recount with Anthropic's token-counting endpoint before assuming the sticker price tells the whole story.

Separately, Anthropic's launch materials say Sonnet 5.5 can cost up to 30% less per task than its immediate predecessor, Sonnet 5, because it needs fewer output tokens to reach a finished answer. That efficiency claim is measured against Sonnet 5, not Sonnet 4.5, so don't double-count it against the list-price cut above.

Context window and output limits

This is the single biggest capability jump between the two models. Claude Sonnet 4.5 has a 200,000-token context window and a 64,000-token maximum output on the synchronous Messages API. Claude Sonnet 5.5 has a 1-million-token context window — roughly 555,000 words on the newer tokenizer — and a 128,000-token maximum output, doubled again to 300,000 tokens on the Message Batches API with a beta header.

On Sonnet 4.5, reaching beyond the default window required a context-window beta header. Sonnet 5.5 ships the 1M window natively, so if your code sets that header today, Anthropic's migration guide says to remove it — sending it alongside Sonnet 5.5 isn't just unnecessary, it's one of the cleanup items on the official migration checklist.

Benchmarks: what Anthropic has published, and what it hasn't

Here's where honesty matters more than a clean headline number. Anthropic's Sonnet 5.5 launch materials publish benchmark comparisons against Claude Sonnet 5, Claude Opus 5.5, and a competing model — not against Claude Sonnet 4.5 directly. The published figures include a Terminal-Bench 4.0 agentic-coding score of 70.6% for Sonnet 5.5 versus 10.3% for Sonnet 5, a Humanity's Last Exam (with tools) score of 64.5% versus 54.9% for Sonnet 5, and an OSWorld 2.1 score of 80.1% versus 57.0% for Sonnet 5. Sonnet 5.5 also scores close to Opus 5.5 on several of these evaluations despite its lower price.

What Anthropic has not published, at least not in any primary source we could find, is a side-by-side Sonnet-4.5-versus-Sonnet-5.5 benchmark table. If you need a quantified capability jump specifically from the model you're retiring, the honest answer is that Anthropic hasn't given you one directly — the comparisons it publishes use Sonnet 5 as the baseline, a model that itself launched after Sonnet 4.5 and before Sonnet 5.5. Treat any "X% better than Sonnet 4.5" claim you see elsewhere as a third-party estimate unless it links back to an Anthropic source, and budget time for your own evals on your actual task rather than relying on general benchmark deltas.

Speed

Anthropic describes Sonnet 5.5 as "our fastest Sonnet model to date," with output generation more than 30% faster than Sonnet 5. Sonnet 4.5 predates that comparison entirely, so there's no official percentage for Sonnet 4.5 versus Sonnet 5.5 specifically. What is documented structurally: Sonnet 4.5 runs with manual "extended thinking," off unless you explicitly enable a thinking budget, while Sonnet 5.5 runs with adaptive thinking on by default, calibrated through an effort parameter that Sonnet 4.5 never had. For latency-sensitive chat or classification work, that default-on reasoning step is worth testing explicitly — you can turn it off by sending thinking: {"type": "between_tools"} at effort levels of high or below, which is the new, low-latency equivalent of running Sonnet 4.5 without a thinking budget.

Availability: API, Bedrock, Google Cloud, Microsoft Foundry, and the apps

Both models are available on the same six Anthropic-tracked surfaces: the Claude API, Amazon Bedrock, Claude Models on Google Cloud (both the Google-managed and Anthropic-managed variants), Microsoft Foundry, and Claude Platform on AWS. The model IDs differ slightly by platform. On the Claude API, Sonnet 4.5 is claude-sonnet-4-5-20250929 (alias claude-sonnet-4-5); Sonnet 5.5 has no dated suffix at all and is simply claude-sonnet-5-5, reflecting Anthropic's newer practice of using dateless, pinned model IDs from the 4.6 generation onward. On Amazon Bedrock's InvokeModel integration, Sonnet 4.5 is anthropic.claude-sonnet-4-5-20250929-v1:0; Sonnet 5.5 on Bedrock is anthropic.claude-sonnet-5-5, served through Bedrock's newer Messages-API-style endpoint rather than the legacy InvokeModel path.

In the Claude consumer apps, you'll simply see Sonnet 5.5 offered as the current Sonnet option once Sonnet 4.5 is retired from a given surface; there's no separate app-side toggle to preserve the old model past its retirement date. For the fuller rundown of where Sonnet 5.5 is offered and what it costs outside the 4.5 comparison, see our explainer on Sonnet 5.5's pricing, speed, and availability.

Sonnet 4.5 vs Sonnet 5.5 at a glance

Migration considerations: what breaks when you switch

Anthropic's own migration guide lists exactly what changes for anyone moving off Sonnet 4.5 specifically (as opposed to a newer model like Sonnet 5). The practical checklist:

• Assistant message prefill is rejected. Sonnet 4.5 accepted a prefilled final assistant turn; Sonnet 5.5 returns a 400 error and requires the conversation to end on a user message. Anthropic's suggested replacements depend on what the prefill was doing — structured outputs for forced formatting, a direct system-prompt instruction for preambles, or moving continuation text into the user turn.
• Forced tool choice no longer works. A tool_choice of type tool or any, both accepted on Sonnet 4.5, now returns a 400 error. The replacement is tool_choice: {"type": "auto"} combined with strict: true on the tool definition, with the prompt telling the model when to use it.
• Set the effort parameter explicitly. Sonnet 4.5 had no effort parameter at all. On Sonnet 5.5, thinking runs by default on any request that omits the thinking field, so you'll want to pick an effort level (Anthropic suggests starting at medium for agentic or latency-sensitive work, high otherwise) rather than inherit whatever the default produces.
• Remove the context-window beta header if your Sonnet 4.5 code sends one — Sonnet 5.5's 1M window is native.
• Computer use needs a new tool version. Sonnet 4.5 used computer_20250124; Sonnet 5.5 rejects it outright and requires the computer_toolset_20260801 toolset on the Claude API and Google-managed Google Cloud (Amazon Bedrock keeps its own computer_20251124 path).
• Beta headers change. Drop interleaved-thinking-2025-05-14 — adaptive thinking interleaves automatically now — and replace fine-grained-tool-streaming-2025-05-14 with an eager_input_streaming: true flag on individual tools.
• Structured outputs move. The old output_format parameter is deprecated on Sonnet 5.5; use output_config.format instead, or add the structured-outputs-2025-11-13 beta header if you must keep the old parameter temporarily.
• Images cost more tokens. Sonnet 5.5 uses a higher-resolution image tier (up to 2,576 pixels on the long edge, versus 1,568 on Sonnet 4.5), which Anthropic says can cost roughly 2.5 times as many tokens for a 2000×1500 image.
• Safety classifiers are stricter. Sonnet 5.5 declines in more categories than earlier Sonnet models and can return stop_reason: "refusal" with categories like cyber, bio, or frontier_llm — worth handling explicitly if your application streams responses to end users.

Anthropic bundles a Claude Code skill (/claude-api migrate this project to claude-sonnet-5-5) that applies the model-ID swap and most of the breaking-parameter changes automatically, then produces a manual checklist — useful if you're migrating a nontrivial codebase rather than a single script.

Should you switch now, or wait?

If you're already past the deprecation notice and have nearly five weeks of runway before November 30, there isn't much reason to wait. Sonnet 5.5 is cheaper per token on paper, has a materially larger context window, and is the model Anthropic itself will keep supporting the longest of the two — its own retirement commitment runs "not sooner than September 28, 2027." The friction is concentrated in a short, well-documented list of breaking changes rather than a vague "things might differ" warning, which makes testing tractable: run your existing eval set against claude-sonnet-5-5, fix the five breaking-change categories above, re-run your effort sweep, and re-check token counts given the roughly 30% tokenizer difference before you lock in a cost estimate.

One thing worth flagging if your workload is specifically latency- or cost-sensitive rather than capability-sensitive: Sonnet 5 — the model released between Sonnet 4.5 and Sonnet 5.5 — is also still an active, non-deprecated option with its own pricing and specs. It's a different comparison than the one in this piece, but worth knowing it exists if "Sonnet 4.5 vs Sonnet 5" rather than "vs Sonnet 5.5" is your actual question; Anthropic's documented replacement path for Sonnet 4.5, in any case, points to Sonnet 5.5 specifically.

Bottom line

Claude Sonnet 4.5 stops working on Anthropic's own platforms after November 30, 2026, and Claude Sonnet 5.5 is the model Anthropic tells you to move to. By the numbers, that's a 33% list-price cut on both input and output tokens, a 5x larger context window, double the maximum output, and a faster model by Anthropic's own description — offset by a tokenizer that uses about 30% more tokens for the same text and five request patterns (prefill, forced tool choice, the old computer-use tool, certain beta headers, and the old structured-output parameter) that now return errors instead of working quietly. None of those are showstoppers, but none of them are "just change the model string," either. Budget a real migration pass, not a find-and-replace, and you'll be off Sonnet 4.5 well before the deadline. For everything new in the replacement model beyond this specific migration, see what's new in Claude Sonnet 5.5.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage