After Claude Haiku 5.5, Why Is the First JSON Block Not the Answer? From budget_tokens to content.type

As of 9 October 2026: Haiku 5.5 launched on 7 October as claude-haiku-5-5. Adaptive thinking is on by default, so the first block can be thinking. budget_tokens returns 400. Read text by content.type, then JSON.parse.

Haiku 5.5 changes which block in the response JSON is the answer. On 7 October 2026 Anthropic released Claude Haiku 5.5. The model ID is claude-haiku-5-5. Adaptive thinking is on by default, so the first item in content can be a thinking block, and by default that block carries only a signature with an empty body. A request that still sends budget_tokens returns 400. Select the text block by type, then hand that string to JSON.parse.

This piece is written for 9 October 2026, from the model page, the migration guide, and the 7 October release notes that were current that day. Our 29 September article on Sonnet 5.5 and between_tools said Haiku 5.5 was still “in the coming weeks.” That date has arrived. This article does not rewrite Sonnet’s four 400s. It only answers what JSON-extraction code has to change from Haiku 4.5 to 5.5.

What shipped on 7 October

The official job is high-volume, latency-sensitive work: classification, routing, extraction, and subagents. The context window is 1M tokens. Max output is 128k tokens. Thinking is adaptive, and the default effort is medium. Platforms are the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Retirement is not sooner than 7 October 2027. claude-haiku-5-5 has no date suffix and no separate alias.

PlatformHaiku 4.5Haiku 5.5
Claude APIclaude-haiku-4-5claude-haiku-5-5
Amazon Bedrockanthropic.claude-haiku-4-5anthropic.claude-haiku-5-5
Google Cloudclaude-haiku-4-5@20251001claude-haiku-5-5
Microsoft Foundryclaude-haiku-4-5claude-haiku-5-5

Prices split on whether the prompt is over 100,000 tokens (USD per million tokens). At or under 100k: input 0.10, output 0.50, cache read 0.01, 5-minute cache write 0.125, 1-hour cache write 0.20. Over 100k: input 0.50, output 2.50, cache read 0.05, 5-minute write 0.625, 1-hour write 1. The Batch API discounts input and output by 50%. The comparison table lists Haiku 4.5 at input 1 and output 5. The launch note also says it costs around 75% less to run on average. The same day, Sonnet 5.5 cache reads dropped from 0.20 to 0.10. That is a different price list. Do not mix it with Haiku’s 0.01 cache read.

The first block is not the answer

Adaptive thinking is on by default. A request that omits thinking can still open with one or more thinking blocks. By default each block’s thinking field is an empty string and only a signature is set. Haiku 4.5 returned summarized thinking. To get a summary, set thinking to adaptive and display to summarized.

The migration guide is direct: select blocks by type, not by position. When you send tool results back, pass thinking blocks through unmodified, including the signature. Thinking tokens count toward max_tokens. A cap tuned for a short Haiku 4.5 reply can stop with stop_reason: "max_tokens" after a thinking block and before any text. Extraction code then feeds an empty or partial thinking block to JSON.parse and reports that the model produced nothing.

{
  "model": "claude-haiku-5-5",
  "max_tokens": 16000,
  "thinking": { "type": "adaptive" },
  "output_config": { "effort": "medium" },
  "messages": [{ "role": "user", "content": "..." }]
}

A forced tool_choice (any, or a named tool) is still accepted on Haiku 5.5. That differs from Opus 5.5 and Sonnet 5.5, where any and tool return 400. On Haiku, a forced call starts the response at the tool call, with no thinking block. To let the model think first, use tool_choice type auto.

Four old JSON shapes that 400

  1. A manual thinking budget. thinking set to enabled with budget_tokens returns 400. Change it to adaptive, and set depth with output_config.effort. Requests that ran on 4.5 with thinking off, or with a tiny budget, take a lower effort here. A simple request can skip thinking entirely.
  2. Sampling parameters. A non-default temperature, a top_p other than 0.99, any top_k, and a request that sends both temperature and top_p, all return 400. If temperature is present it must be 1. If top_p is present it must be 0.99; 1 is rejected. Delete all three fields.
  3. Assistant prefill. Ending messages on an assistant turn, so the model continues it, returns 400 even with thinking off. Move format constraints to structured outputs, or to tools with enum fields for classification. Put continuations in a user message.
  4. The old computer-use tool. On the Claude API and Google Cloud, computer_20250124 plus the old beta header returns 400. Replace it with the computer_toolset_20260801 toolset and drop computer-use-2025-01-24. Sending fine-grained-tool-streaming-2025-05-14 beside a toolset also returns 400.

One failure is not a 400, and it still drops the reasoning: replay a Haiku 5.5 thinking block through a different account and the API removes the block before the model sees it. The request succeeds. A multi-tenant store has to replay each conversation through the account that produced it. Changing system, tools, or earlier messages and then sending the thinking block back returns 400. Keep the conversation append-only. On accounts created before 31 August 2026, 00:00 UTC, that error fires only when the request sets prefix_mismatch_behavior.

Handle stop_reason: "refusal" on its own. Haiku 5.5 runs safety classifiers and has no server-side fallback. Priority Tier is not supported on this model. For Claude Managed Agents, the migration guide says no change beyond the model name.

Thinking off, plus structured output

thinking type disabled works at low, medium, and high. At xhigh and max the request returns 400. Telling the model in the prompt to answer directly did not stop thinking in Anthropic’s tests. Lower effort is the lever for less thinking.

With thinking off and structured outputs requesting JSON, the model can skip a tool it needed. Three paths: switch back to adaptive thinking; drop output_config.format on any request that must call a tool; or force the call with tool_choice. A forced call writes no text before the call. The prompting guide suggests a system line: the JSON format applies to the final answer only; when a tool is needed, call it first, then write the JSON after the result.

Claude on Amazon Bedrock does not support structured outputs. Use tools with enum fields for classification and extraction. Do not wait for output_config.format there. Changing the top-level effort between turns invalidates the prompt cache for the conversation. A per-message effort change needs the mid-conversation-output-config-2026-07-01 beta header and adaptive thinking. With thinking off, that mid-conversation change itself returns 400.

The same text counts as more tokens

Haiku 5.5 uses the newer tokenizer shared with Claude 4.7 and later models. The same input counts as approximately 30% more tokens than on Haiku 4.5. The exact increase depends on the content. Requests, responses, and streaming events keep the same shape. What changes is anything you budget in tokens: usage is higher, the same max_tokens holds less text, and cost estimates built from 4.5 counts have to be recomputed. Count prompts with model set to claude-haiku-5-5. Do not reuse 4.5 counts.

The “around 75% less to run” line and the “about 30% more tokens” line belong on the same sheet. Swapping only the model string, while keeping the old max_tokens and the old token counts, cuts extraction off before the JSON is finished.

Four JSON documents to check first

  1. The request body. No budget_tokens, no non-default sampling, and messages does not end on an assistant turn. Thinking is omitted or set to adaptive.
  2. One real response. Take the block whose type is text, then JSON.parse it. Do not parse content[0].
  3. The thinking block you send back. The signature is still there, the block was not rewritten, and system and tools are unchanged from the turn that produced it.
  4. The structured-output schema. If this turn must call a tool, do not also turn thinking off. On Bedrock, use a tool enum instead of output_config.format.

Read the response with local JSON tools

Before you change the model string, open three texts in the browser: the old Haiku 4.5 request, the new 5.5 request, and one response that contains a thinking block.

  • JSON Validator — whether the request and the response are legal JSON, and whether a schema still requires a field you deleted.
  • JSONPath — $..content[?(@.type=='text')] pulls the answer and skips thinking.
  • JSON Diff — compare the 4.5 request with the 5.5 request and confirm budget_tokens, temperature, and the prefill are gone.

Nothing leaves the browser. Read the contract flat, then change the model ID. Haiku 5.5 decides how long to think. Your parser should not treat the first empty thinking block as the extraction result.

FAQ

Can I ship by only changing the model to claude-haiku-5-5?

Not as a default. budget_tokens, a non-default temperature or top_p, any top_k, an assistant prefill, or computer_20250124 on the Claude API or Google Cloud, all return 400. A parser that still reads the first content block will treat thinking as the answer.

Why is the first parsed block empty?

By default a thinking block’s thinking field is an empty string and only the signature is set. Summarized text needs display set to summarized. The answer is the block whose type is text, not index 0.

Can a smaller budget_tokens cut the bill?

No. Enabled thinking that includes budget_tokens returns 400. Less thinking means a lower effort. A simple request at low effort can produce no thinking block.

If I turn thinking off and still ask for JSON, will tools still be called?

Not reliably. The prompting guide says that with thinking off and structured outputs, the model can skip a tool it needs. Keep adaptive thinking, drop output_config.format, or force the call with tool_choice.

How do I get structured JSON on Bedrock?

The migration guide says Claude on Amazon Bedrock does not support structured outputs. Use tools with enum fields for classification and extraction.

Is this the same rule as September’s Sonnet 5.5 between_tools change?

No. Sonnet 5.5 rejects thinking.disabled and wants between_tools, and it rejects a forced tool_choice. Haiku 5.5 allows disabled at high and below, and it accepts a forced tool_choice, but that response has no thinking block.

Summary

Haiku 5.5 retires the assumption that the first block is the JSON answer. It does not retire extraction. claude-haiku-5-5, launched 7 October, thinks adaptively by default, rejects budget_tokens, and counts the same text as about 30% more tokens. Read text by content.type, pass thinking blocks back unchanged, and do not bind structured output to thinking-off on a turn that must call a tool.

Flatten the old request, the new request, and one real response locally before you change the model ID. The price list can be lower. The parser still has to know which block is the JSON.