Claude Sonnet 4.6 vs Claude Sonnet 5.5: What Migrating Changes

A team still running Claude Sonnet 4.6 in production today would not move to Claude Sonnet 5, the model this guide first compared it with in August. It would move to Claude Sonnet 5.5, which Anthropic released on September 28, 2026 at $2 per million input tokens and $10 per million output tokens. The older model has not gone anywhere: Anthropic's deprecations page still lists claude-sonnet-4-6 as Active, with retirement not sooner than February 17, 2027.

So the question has the same shape as it did in August, with a different destination. Nothing forces the move, the swap is a one-line model ID change, and the work that decides whether it goes smoothly sits in the request parameters, the tokenizer and the thinking settings. This page re-read Anthropic's models overview, deprecations page, the Sonnet 5.5 migration guide and the Claude Code model documentation on October 9, 2026, and everything below is checked against them.

What changed since the August 4 version

Four things moved, and one claim in the first version of this guide no longer holds.

• The destination moved. Sonnet 5.5 shipped on September 28, 2026. Anthropic's migration guide now lists Sonnet 4.6 as a starting point for the move to 5.5, with its own checklist.
• The price story changed. The first version warned that Sonnet 5's introductory price would end on August 31 and that the bill would jump on September 1. Anthropic's current pages show Sonnet 5.5 at $2 and $10 per million tokens and say its prices match Sonnet 5, apart from cheaper cache reads. That September 1 warning is withdrawn.
• Claude Code changed its default. The first version said Sonnet 5 was the default for Pro and Team Standard seats. Claude Code's documentation now says the default resolves to Claude Opus 5.5 across its plans.
• Sonnet 4.6 stayed put. It is still Active and has no retirement notice, which is the part of the August verdict that held.

Re-checked October 9, 2026: the central advice holds, and only the target model is different.

The verdict: migrate the request, not the workflow

For a team calling the API with default settings, moving from Sonnet 4.6 to Sonnet 5.5 is a model ID change plus a short list of parameters to remove. For a team with tuned token budgets, sampling controls, forced tool choice or a reasoning-budget pipeline, it is a small engineering project.

The reasons to move are real but not urgent. Sonnet 5.5 costs less per token than the $3 and $15 that Sonnet 4.6 was priced at, it lists a June 2026 knowledge cutoff and it is the model Claude Code's sonnet alias now points to. The reasons to wait are also real: the tokenizer produces more tokens for the same text, and five request settings that Sonnet 4.6 accepts now return errors.

What it costs

The model page for Sonnet 5.5 shows a 1 million token context window and 128K tokens of maximum output. Cache reads cost $0.10 per million tokens, cache writes are priced separately for five-minute and one-hour holds, and the Batch API takes 50% off.

Sonnet 4.6 no longer appears in the overview's comparison table, so its price here comes from the earlier version of this guide, checked on August 4, 2026: $3 input and $15 output per million tokens, with a 1 million token window and 128K output.

Anthropic's models overview, captured October 9, 2026: Sonnet 5.5 sits at $2 and $10 per million tokens.

On list price alone the newer model is a third cheaper.

Whether the invoice follows depends on the tokenizer. Anthropic states that against Sonnet 4.6 the same text produces about 30% more tokens on Sonnet 5.5, depending on the content. A third off the rate and thirty percent more tokens roughly cancel, so a team should measure instead of assuming a saving. Images matter too: Sonnet 5.5 uses the high-resolution image tier, and Anthropic's own example is a 2000 by 1500 pixel image costing about 2.5 times as many tokens as on Sonnet 4.6.

The cheapest way to find out is to run a week of real prompts through both models and compare token counts and dollars. The token-counting API gives the count without running the request. For a wider view of where Claude models sit on price against other labs, the model leaderboard shows current figures, and the recommendation tool filters by budget.

A worked example of the cost math

Take an illustrative workload of 100 million input tokens and 20 million output tokens a month on Sonnet 4.6, billed at the rates in the table above. Move the same text to Sonnet 5.5 and apply Anthropic's roughly 30% token increase, which gives 130 million input and 26 million output tokens. The bills below follow from those counts.

The saving in this example is about 13%, not the one third that the price cut suggests. The gap is the tokenizer. A workload heavy on images would land lower, because of the high-resolution tier, and a workload that leans on cached prompts would land higher, because cache reads cost $0.10. The only honest number is the one from your own traffic.

Where Sonnet 5.5 wins

The case for moving is strongest in four places.

• Price per token. $2 and $10 against $3 and $15, before caching. Cache reads at $0.10 per million tokens make long, repeated prompts cheaper still.
• Lifespan. Retirement is not sooner than September 28, 2027, seven months after Sonnet 4.6's earliest date. A team that migrates now does not have to do it again soon.
• Capability. Anthropic describes Sonnet 5.5 as the best combination of speed and intelligence in its current line. For benchmark figures, Claude Sonnet 5 is the cleaner reference: HokAI's records show 82.1% on SWE-bench Verified for Sonnet 5 against 79.6% for Sonnet 4.6, and the 5.5 record carries its own sourced scores.
• Tooling. Anthropic's migration guide notes that the bundled Claude API skill can apply the model swap and the breaking parameter changes for you. In Claude Code, running /claude-api migrate this project to claude-sonnet-5-5 starts it, and it asks you to confirm the scope before editing files.

None of these is a reason to rush. They are reasons to schedule it.

Where Sonnet 4.6 still wins

The strongest counterargument is that nothing requires anyone to move. Sonnet 4.6 is Active, and the deprecations page gives it a retirement date of not sooner than February 17, 2027. Compare that with a model on a hard clock: the deprecations page lists claude-sonnet-4-5-20250929 for retirement on November 30, 2026, with claude-sonnet-5-5 as its named replacement. Sonnet 4.6 has no such notice.

Four situations favour staying on the older model for now.

1. Your code sets sampling parameters. Sonnet 4.6 accepts temperature, top_p and top_k. On Sonnet 5.5 a non-default value returns a 400 error.
2. Your code uses manual thinking budgets. Sonnet 4.6 still accepts thinking.type: "enabled" with a budget_tokens value, which Anthropic marks deprecated. Sonnet 5.5 accepts only adaptive and between_tools.
3. Your pipeline forces a tool. A tool_choice of type any or tool works on Sonnet 4.6 and returns a 400 error on Sonnet 5.5, including on the token-counting endpoint.
4. Your cost model is built on old token counts. A budget that was tight on Sonnet 4.6's tokenizer needs rebuilding before it holds on the new one, and rebuilding on a schedule beats rebuilding under deadline pressure.

The word "legacy" on a documentation page describes product strategy. It is not an operational deadline, and a stable integration with no cost or capability problem can reasonably wait for the real date.

What switching takes

Anthropic's migration guide gives a checklist for anyone starting from Sonnet 4.6 or earlier. The practical version has two parts: what every starting model must do, and what Sonnet 4.6 users must add.

For every starting model:

1. Change the model ID to claude-sonnet-5-5.
2. Read response content blocks by type and pass thinking blocks back unchanged in tool loops.
3. Replace forced tool use with auto and strict tools.
4. Keep conversations append-only, because thinking blocks are tied to the model that produced them.
5. On the Claude API and Google Cloud, move computer use to the newer toolset, since computer_20251124 is not accepted.
6. Handle refusals and configure fallback.
7. Re-run your effort sweep and re-baseline cost.

Added for Sonnet 4.6 and earlier:

1. Expect thinking on a request that has no thinking field. On Sonnet 4.6 that request ran without it, and on 5.5 it runs with adaptive thinking.
2. Replace thinking budgets with an effort level.
3. Remove temperature, top_p and top_k.
4. Recount tokens, revisit max_tokens and plan for the high-resolution image tier.

Thinking deserves its own warning. On Sonnet 5.5 thinking text is omitted by default, so thinking blocks arrive with an empty text field and a signature. Code that reads content[0].text breaks, because a response can begin with a thinking block. Anthropic says max_tokens now covers thinking plus text, and thinking tokens bill as output. To keep running without up-front thinking, send thinking: {"type": "between_tools"} at high effort or below. The setting Sonnet 5 used, disabled, returns a 400 error on 5.5.

Prefill is the one item on the list that is not new for this move. Sonnet 4.6 already rejects an assistant prefill, so a team that got that far on 4.6 has done it.

Data handling, security and refusals

Anthropic's migration guide does not describe any change to data retention terms for the move, and this guide does not claim one. What it does describe is a change in behaviour that matters for security reviews.

Sonnet 5.5 declines in more categories than Sonnet 5. A declined request returns stop_reason: "refusal", and stop_details can name a category such as cyber or bio. A security team that runs red-team tooling, malware analysis or exploit research through the API should test those workloads before cutover. The migration checklist lists refusal handling and fallback configuration as a required step for every starting model.

Two smaller points. Because thinking text is omitted by default, application logs that used to hold a reasoning summary will hold an empty field unless you set display: "summarized". And because thinking blocks are bound to the model and conversation, a stored transcript cannot be replayed against a different model without breaking the signature.

Where Claude Code users already felt this

For Claude Code users, the choice has mostly been made for them. Claude Code's documentation says the sonnet alias resolves to Sonnet 5.5 on the Anthropic API, and that Sonnet 5.5 needs Claude Code version 2.1.284 or later. The default setting now points at the Opus tier for every paid plan and for the Anthropic API. Before version 2.1.280 it resolved to Sonnet 5 on Pro and Team Standard, which is the claim the first version of this guide made.

Provider matters. The same documentation shows the sonnet alias resolving to Sonnet 4.6 on Claude Platform on AWS, and to Sonnet 4.5 on Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry. A team on those providers is not on 5.5 unless it pins the full model name. Claude Fable 5.1 sits above all of these and is not the default on any plan.

For a wider comparison of the whole line, the guide on which Claude model to use covers Claude Opus 5.5, Sonnet and Claude Haiku 5.5, and the Haiku 5.5 versus Haiku 4.5 guide covers the same kind of move one tier down. For how Claude Code compares with other coding agents, see Cursor Composer 2 versus Claude Code.

A test plan that fits in an afternoon

Treat the move as an experiment with a clear pass mark.

1. Copy a week of production prompts into a test rig.
2. Count tokens for each prompt with the token-counting API on both models, and record the ratio.
3. Run the prompts on Sonnet 5.5 with the parameters removed, and compare output quality on the cases your users complain about most.
4. Check the error log for 400 responses, which flag every leftover sampling or forced-tool setting.
5. Multiply the measured token ratio by the new rates and compare against your current bill.

If the quality holds and the bill drops, cut over gradually. If either fails, the test cost an afternoon and the integration stays where it is. Teams that already route work between coding assistants such as Cursor or GitHub Copilot can run the same rig against the model setting in each tool. Heavy users who hit plan limits first should read how Claude Max users run out by Wednesday, because token growth shows up there before it shows up on an invoice.

Who should pick which

• Pick Sonnet 5.5 if you are starting a new build, you call the API with default settings, or you want the longest runway before the next forced migration.
• Stay on Sonnet 4.6 if you depend on sampling parameters, thinking budgets or forced tool choice, and the rewrite is not scheduled yet.
• Test both if your volume is high enough that a 30% token difference moves your invoice. Run real prompts, not benchmarks.
• Skip both if your workload is classification, extraction or routing. Haiku 5.5 is the line's fast tier, and the guide to choosing the right LLM walks through the trade-off.

Teams that are unsure which tier fits can answer a few questions in Smart Match and get a shortlist matched to their use. Readers still deciding whether Claude is the right family at all will find what Claude is for and the comparison of ChatGPT, Claude, Copilot, Grok and Poe useful.

What would change the verdict

Three developments would flip the advice in this guide.

• A retirement notice for Sonnet 4.6. If Anthropic moves claude-sonnet-4-6 from Active to deprecated, "wait" becomes "migrate before the date". Watch the deprecations page, which is where the claude-sonnet-4-5-20250929 notice appeared.
• A measured bill that goes up. If a week of your own prompts costs more on 5.5 than on 4.6 because of the tokenizer and image tier, the lower list price does not pay for the move, and the case for staying strengthens.
• A newer Sonnet. Sonnet 5 followed 4.6 within months, and 5.5 followed 5 within three. A team planning a migration this quarter should check the overview page the day before it starts, because the target may have moved again. The Anthropic company page lists the current line, and the Opus 5 and Opus 4.8 records show how quickly earlier flagships were superseded.

If none of these happens, the safest plan is the dull one. Test on a copy of production traffic this month, fix the parameters that error, and cut over when the numbers say it pays.

Keep the rollback simple by leaving the old model ID in configuration for a few weeks after the switch, so a surprise in output quality or cost can be reversed in minutes. Log the token counts per request during that window, because the first month of real data is the only evidence that settles the question for your own workload.

Frequently asked questions

Is Claude Sonnet 4.6 deprecated?

No. Anthropic's deprecations page lists claude-sonnet-4-6 as Active, with retirement not sooner than February 17, 2027. Claude Sonnet 4.5 is the one on a clock: Anthropic announced its retirement for November 30, 2026, naming Sonnet 5.5 as the replacement.

What is the difference between Claude Sonnet 4.6 and Claude Sonnet 5.5?

Sonnet 5.5 costs $2 and $10 per million tokens against $3 and $15, runs adaptive thinking by default and uses a tokenizer that produces about 30% more tokens. It rejects temperature, top_p, top_k, forced tool choice and thinking budgets, all of which Sonnet 4.6 accepts except forced tool use being supported there too.

Is Sonnet 5.5 the default model in Claude Code?

No. Claude Code's documentation says the default resolves to Opus 5.5 on Pro, Max, Team, Enterprise and the Anthropic API, while the sonnet alias resolves to Sonnet 5.5 on the Anthropic API. Sonnet 5.5 needs Claude Code 2.1.284 or later.

Does migrating to Claude Sonnet 5.5 cost more?

Not on list price, which falls from $3 and $15 to $2 and $10 per million tokens. But the same text produces about 30% more tokens, and images use a high-resolution tier that can cost about 2.5 times as many tokens, so measure a week of real prompts before assuming a saving.

Should I migrate from Claude Sonnet 4.6 to Claude Sonnet 5.5 right now?

If your integration uses default settings, yes, it is close to a one-line change. If you set sampling parameters, thinking budgets or forced tool choice, schedule the rewrite and test on real traffic first, since Sonnet 4.6 has no retirement notice.

Covered in this guide

• Claude Sonnet 4.6: Claude Sonnet 4.6 by Anthropic (Feb 2026) scores 79.6% on SWE-bench Verified with a 1M-token context window, and is now one generation behind Sonnet 5.5.
• Claude Sonnet 5.5: Anthropic's fastest Sonnet-class model, launched in September 2026, with a 1M context and output more than 30% faster than Sonnet 5.
• Anthropic: Anthropic is a Public Benefit Corporation started by 7 ex-OpenAI researchers. It builds the Claude model family and counts Amazon and Google among its largest backers.
• Claude Code: Anthropic's agent for the terminal, IDEs, desktop, web and mobile that edits files, runs tests and opens pull requests. There is no free plan.
• Claude Fable 5.1: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon agentic coding and research at Fable 5's per-token rates.
• Claude Haiku 5.5: Anthropic's smallest and fastest 5.5 model, released October 7, 2026, with a 1M context window, built for high-volume work.
• Claude Opus 4.8: Claude Opus 4.8 (May 2026) is Anthropic's former flagship, now a legacy release with a 1M-token context window. Its newer successor is the recommended choice for new builds.
• Claude Opus 5: Anthropic's July 2026 Opus with a 1M token context and xhigh reasoning effort, still active but superseded by Claude Opus 5.5 on September 22, 2026.
• Claude Opus 5.5: Anthropic's September 2026 flagship LLM with a 1 million token context window, built for long-running agentic coding and computer use.
• Claude Sonnet 5: Claude Sonnet 5, released by Anthropic on June 30, 2026, hits 82.1% SWE-bench Verified with a 1M-token context window.
• Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
• GitHub Copilot: GitHub's AI coding assistant with agent mode and MCP support across 6 IDEs, plus a free plan with no credit card required.

Sources