Claude Sonnet 5.5: Price, What Changed From Sonnet 5, and the 5 Breaking Changes

Claude Sonnet 5.5: Price, What Changed From Sonnet 5, and the 5 Breaking Changes

Claude Sonnet 5.5 came out on September 28, 2026 at the same list price as Sonnet 5: $2 per million input tokens and $10 per million output tokens. The price is the boring part. The part that will page you at 2 a.m. is that Sonnet 5.5 returns a 400 error for two settings a lot of production code sends every day: thinking: {"type": "disabled"} and forced tool use through tool_choice.

If you run classifiers, extractors or agents on claude-sonnet-5, this guide covers what you pay, what actually changed, the five breaking changes Anthropic lists with the fix for each, and when Sonnet 5.5 is the right pick against Opus 5.5 and Haiku 5.5. Every number comes from Anthropic's own model page, pricing page, migration guide and launch post, all read on October 23, 2026.

TL;DR

• Price: $2 / $10 per million tokens, the same as Sonnet 5. Cache reads drop to $0.10. Batch API is 50% off.
• Specs: 1M context, 128K output (300K on the Batch API with a beta header), adaptive thinking on by default, default effort high.
• What Anthropic says improved: 30%+ faster output and "up to 30% less per task" in its own testing, because it uses fewer tokens for the same work.
• What breaks: thinking: disabled, tool_choice of any or tool, edited conversation history that replays thinking blocks, computer_20251124 on the Claude API, and four advisor models.
• What goes quiet: text the model writes between tool calls now arrives in thinking blocks, empty by default.

How much does Claude Sonnet 5.5 cost?

Sonnet 5.5 keeps Sonnet 5's base prices. The one price that moved is the cache read, which Anthropic cut to 5% of the input price instead of the usual 10%.

Three pricing details that are easy to miss:

1. The full 1M context is billed at the standard rate. Anthropic's pricing page says a 900k-token request costs the same per token as a 9k-token one on Claude 4.6 and later models (Haiku 5.5 is the exception).
2. US-only inference costs 1.1x. Setting inference_geo: "us" multiplies every token price, cache reads included.
3. Tool definitions got cheaper to send. The hidden tool-use system prompt is 286 tokens on Sonnet 5.5 against 354 on Sonnet 5 (with tool_choice set to auto or none).

What does that mean for a real workload?

Take a support classifier that runs 10,000 requests a day, each with 3,000 input tokens (2,500 of them a shared system prompt and tool list) and 500 output tokens. This is our worked example on list prices, not a measurement.

The list-price saving is small: $2.50 a day in the cached case. Inference: the bigger saving, if you see one, will come from Sonnet 5.5 producing fewer output tokens per task, which Anthropic and several of its launch testers report. Balyasny Asset Management, quoted in the launch post, measured about 121k tokens per answer on Sonnet 5.5 against 497k on Sonnet 5 across 2,441 finance tasks. Run your own token count before you bank that.

What changed from Sonnet 5?

The model ID, the knowledge cutoff and the defaults around thinking all moved. The tokenizer did not: Anthropic says the same text produces the same token counts on both models, so your token budgets carry over.

How much better is it, according to Anthropic?

These are Anthropic's published scores from the launch post. We have not reproduced them.

Anthropic notes that Artificial Analysis ran GDPval-AA on a pre-release deployment that had a structured-outputs bug, since fixed. The launch post also says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment", so read the near-tie on GDPval as a statement about that benchmark, not about every task.

What are the 5 breaking changes in Sonnet 5.5?

Anthropic's "What's new" page lists five changes that make Sonnet 5 code fail on Sonnet 5.5. Here is each one, the error you will see, and the fix.

1. Why does thinking: disabled return a 400 error?

Sonnet 5.5 has no "off" switch for thinking in the old sense. Send thinking: {"type": "between_tools"} instead. It turns off up-front thinking, and if your request uses no tools, the response is plain text, as disabled was on Sonnet 5.

# Before (Claude Sonnet 5)
client.messages.create(
    model="claude-sonnet-5",
    max_tokens=16000,
    thinking={"type": "disabled"},
    messages=[{"role": "user", "content": "..."}],
)

# After (Claude Sonnet 5.5)
client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=16000,
    thinking={"type": "between_tools"},
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "..."}],
)

Three rules come with it, all from the migration guide:

• between_tools works at low, medium and high effort only. At xhigh or max it returns a 400, so use adaptive thinking there.
• It takes no other fields. Sending display, budget_tokens or block_binding alongside it is a 400.
• Effort can't change mid-conversation while you use it. Manual budgets ({"type": "enabled", "budget_tokens": N}) are also rejected.

2. Why does tool_choice fail on Sonnet 5.5?

Sonnet 5.5 does not support forced tool use. A tool_choice of {"type": "any"} or {"type": "tool", "name": "..."} returns this 400, on the Messages endpoint and on token counting:

tool_choice: type "tool" and "any" are not supported for this model.

This is the one most likely to hit classifier and extraction code, because forcing a single "record_result" tool is a common way to get structured output. The fix Anthropic gives: send tool_choice: {"type": "auto"}, mark the tool strict: true so its input matches the schema, and say in the prompt when the tool applies.

client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    tools=[{**tool, "strict": True} for tool in tools],
    tool_choice={"type": "auto"},
    messages=[{"role": "user", "content": "What's the weather in Paris? Use the get_weather tool."}],
)

With auto, the model can still answer in text instead of calling the tool. If your pipeline needs JSON every time, Anthropic points to structured outputs as the other route. Strict tool use supports a subset of JSON Schema and needs additionalProperties: false on every object, so check your schemas before you flip the flag.

3. What does "thinking blocks are tied to the model and the conversation" mean?

Every Sonnet 5.5 thinking block is signed over everything before it: the system prompt, the tools and earlier messages. For accounts created on or after August 31, 2026, 00:00 UTC, the API enforces that check by default on the Claude API, Amazon Bedrock and Google-managed Google Cloud. Replay a block after editing earlier history and you get a 400.

The fix is to keep conversations append-only. Change instructions or tools with mid-conversation system messages rather than by editing the system field. If you must edit, the thinking-binding-controls-2026-08-01 beta header with block_binding.prefix_mismatch_behavior: "drop_block" drops the stale blocks instead of failing (adaptive thinking only).

Two related limits: Sonnet 5.5 can't read thinking blocks from Opus 5, Opus 5.5 or any Fable or Mythos model (the API drops them, the request still succeeds), and its blocks only work in the account that produced them or a linked one.

4. Why is computer_20251124 rejected?

On the Claude API and on Google-managed Google Cloud, Sonnet 5.5 supports computer use only through the computer_toolset_20260801 toolset. The old tool returns:

'claude-sonnet-5-5' does not support tool types: computer_20251124.

Replace the tools entry with {"type": "computer_toolset_20260801"}, drop the old beta header, and update your agent loop for member tool_use blocks and batched actions. On Amazon Bedrock, computer_20251124 still works. If you send the fine-grained-tool-streaming-2025-05-14 header, remove it when you switch: next to a toolset it is a 400.

5. Which advisor tool pairings stop working?

With the advisor tool (beta), a Sonnet 5.5 executor rejects Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 5.5 as advisors. The migration guide also lists Opus 4.6 and Sonnet 4.6. Accepted advisors are Opus 5, Opus 5.5, Sonnet 5.5 itself, Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1. Every accepted advisor returns its advice encrypted as an advisor_redacted_result block, so your code can't read the advice text.

The silent one: where did the text between tool calls go?

This change fails no request, which makes it easier to miss. On Sonnet 5.5, notes longer than a sentence or two written between tool calls come back as thinking blocks, and at the default display: "omitted" they are empty. An agent UI that streamed "Now checking the invoice table..." goes quiet. With adaptive thinking, set display to "summarized", or to "updates" (beta header thinking-display-updates-2026-08-18). With between_tools, the text comes back without any setting.

What is the Sonnet 5.5 migration checklist?

Anthropic's own list, in the order we would work through it on a client codebase:

1. Search for "disabled" in thinking configs. Replace with between_tools, and check the effort level is high or lower.
2. Search for tool_choice. Replace any and tool with auto plus strict: true, and add a line to the prompt saying when to call the tool.
3. Find any code that edits system, tools or earlier messages mid-conversation. Make it append-only.
4. Move computer_20251124 to computer_toolset_20260801 on the Claude API.
5. Swap rejected advisor models.
6. If your UI shows progress text between tool calls, set thinking.display.
7. Read content blocks by type, not content[0].text, since a response can start with thinking blocks.
8. Re-run your effort sweep. Anthropic says effort levels are recalibrated, so high on Sonnet 5.5 is not high on Sonnet 5. Its starting points: high by default, medium for well-specified agentic coding, medium or low for chat.

One more thing to plan for: Sonnet 5.5 is the first Sonnet launched with cyber safeguards. Declines come back as HTTP 200 with stop_reason: "refusal" and a stop_details category (cyber, bio, frontier_llm, reasoning_extraction or general_harms). Server-side fallback (fallbacks: "default", beta) retries cyber and frontier_llm declines on Sonnet 5. Security tooling should handle that stop reason explicitly.

Should you use Sonnet 5.5, Opus 5.5 or Haiku 5.5?

Anthropic's models overview says to start with Opus 5.5 when you are unsure. Our rule of thumb for client work, using the published prices:

Haiku 5.5's price rises fivefold, to $0.50 / $2.50, for prompts over 100,000 tokens, so long-context jobs usually belong on Sonnet or Opus. Inference: for most teams the split is Haiku for the high-volume first pass, Sonnet 5.5 for the default agent model, and Opus 5.5 for the hard 10%. If you want the longer version of that decision, our Opus vs Sonnet vs Haiku comparison walks through it task by task, and the Claude Opus 5.5 release guide covers Opus 5.5's own breaking changes.

Where can you use Claude Sonnet 5.5?

It is on the Claude API for all customers as claude-sonnet-5-5, on Amazon Bedrock as anthropic.claude-sonnet-5-5, and as claude-sonnet-5-5 on Claude Platform on AWS, both Google Cloud options (Anthropic managed and Google managed) and Microsoft Foundry. Anthropic says it is also available with zero data retention. In Claude Code and the Claude apps the default effort is medium; on the Claude Platform it is high.

FAQ

What is Claude Sonnet 5.5?

Claude Sonnet 5.5 is Anthropic's mid-tier model in the Claude 5.5 family, released September 28, 2026, with the API ID claude-sonnet-5-5. Anthropic describes it as "the best combination of speed and intelligence" and positions it as a faster, cheaper complement to Opus 5.5 for well-scoped everyday tasks.

How much does Sonnet 5.5 cost per million tokens?

Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. Cache reads are $0.10, 5-minute cache writes $2.50, 1-hour cache writes $4, and the Batch API halves input and output to $1 and $5. US-only inference adds a 1.1x multiplier.

Is Sonnet 5.5 cheaper than Sonnet 5?

Sonnet 5.5 has the same list price as Sonnet 5 except cache reads, which are half price ($0.10 against $0.20). Anthropic says it costs "up to 30% less per task" in its testing because it uses fewer tokens, but that depends on your workload, so measure output tokens per task on your own traffic.

Why do I get "tool_choice: type tool and any are not supported for this model"?

You are sending forced tool use to Sonnet 5.5, which does not support it. Change tool_choice to {"type": "auto"}, set strict: true on the tool, and tell the model in the prompt when to call it. For guaranteed JSON, use structured outputs instead.

How do I turn off thinking on Sonnet 5.5?

Send thinking: {"type": "between_tools"}. The old {"type": "disabled"} returns a 400 error on Sonnet 5.5. between_tools works at low, medium or high effort and takes no other thinking fields.

Is Sonnet 5 being retired?

Not yet. Anthropic lists Sonnet 5 as "Active (legacy)" with retirement not sooner than June 30, 2027, and recommends migrating to Sonnet 5.5. Sonnet 5.5's own retirement date is not sooner than September 28, 2027.

What is the context window of Sonnet 5.5?

Sonnet 5.5 has a 1M token context window and up to 128K output tokens on the Messages API, or 300K on the Message Batches API with the output-300k-2026-03-24 beta header. The full window is billed at standard rates.

Should I switch from Sonnet 5 to Sonnet 5.5 now?

Switch once your code passes the five breaking changes and your evals hold. The model ID swap alone is not enough if you use disabled thinking, forced tool_choice, edited conversation history, computer_20251124 or an old advisor model. If you want help auditing a production codebase for those, our Claude Code development team does this kind of migration.

---

Sources: Claude Sonnet 5.5 model page, What's new in Claude Sonnet 5.5, Claude Sonnet 5.5 migration guide, Claude pricing, Claude Sonnet 5 model page, Claude Haiku 5.5 model page, Claude models overview, Introducing Claude Sonnet 5.5 (Anthropic). All read October 23, 2026.