Claude Haiku 5.5 starts at $0.10 per million tokens, and five Haiku 4.5 habits now return a 400

Claude Haiku 5.5 is cheap per token, counts about 30% more tokens for the same text, and turns temperature, prefill and manual thinking budgets into HTTP 400 errors.

Anthropic released Claude Haiku 5.5 on October 7. The release note describes it as "our most capable model tuned for high-volume and latency-sensitive work," and the model overview names the target jobs: classification, routing, extraction and subagent tasks.

The specifications: a 1M-token context window (up from 200k on Haiku 4.5), 128k output tokens (up from 64k), adaptive thinking with a default effort of medium, and a June 2026 knowledge cutoff. The ID is claude-haiku-5-5, which the migration guide calls "a fixed model ID with no date suffix and no separate alias." Retirement is "Not sooner than October 7, 2027."

The price, and the tokenizer under it

The overview's pricing table, per million tokens:

	Prompts up to 100,000 tokens	Prompts over 100,000 tokens
Input	$0.10	$0.50
Output	$0.50	$2.50
Cache read	$0.01	$0.05
5-minute cache write	$0.125	$0.625
1-hour cache write	$0.20	$1

Batch requests are 50% off. Two things move the real bill. First, every rate rises fivefold once a prompt passes 100,000 tokens. Second, the tokenizer: "the same input text produces approximately 30% more tokens on Claude Haiku 5.5 than on Claude Haiku 4.5." Images cost more too, because Haiku 5.5 uses the high-resolution tier; the guide's example is a 2,000 by 1,500 pixel image costing "about 2.5 times as many visual tokens." One change runs the other way: the minimum cacheable prompt drops to 512 tokens, from 4,096 on Haiku 4.5.

Five ways to get a 400

The release note is blunt: "Code written for Claude Haiku 4.5 can break on Claude Haiku 5.5." The migration guide lists what now fails:

Manual thinking budgets. {"type": "enabled", "budget_tokens": N} returns a 400. Use {"type": "adaptive"} and steer with output_config.effort.
Sampling parameters. temperature must be 1 if sent, and top_p must be 0.99. Any other value errors, "including a top_p of 1," as does any top_k, and any request carrying both temperature and top_p.
Assistant prefill. A final assistant turn is rejected "even with thinking turned off."
The old computer use tool. computer_20250124 errors everywhere; the replacement is computer_toolset_20260801 (Claude API, Google Cloud) or computer_20251124 (Amazon Bedrock).
Edited history with thinking blocks. Sending a thinking block back after changing system, tools or earlier messages returns a 400. On accounts created before August 31, 2026, 00:00 UTC, this applies only to requests that set thinking.block_binding.prefix_mismatch_behavior.

The thinking change, as the guide shows it:

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

Separately, structured outputs are not available for Haiku 5.5 on Amazon Bedrock, and Priority Tier is not supported on the model at all.

The quieter changes

Several changes do not error but will surprise a parser. Adaptive thinking is on by default, so a response can open with thinking blocks the request never asked for; the guide says to select blocks by type, not position. Those blocks arrive with an empty thinking field unless you set "display": "summarized". Thinking tokens count toward max_tokens, so a tight limit "can stop with stop_reason: \"max_tokens\" after a thinking block and before any text." Thinking blocks from all earlier turns stay in context as input tokens, and blocks replayed through a different account are dropped silently. The model also runs safety classifiers that can return stop_reason: "refusal", with no server-side fallback.

Why a Haiku user should care

Our reading, not Anthropic's: the patterns that break are the ones small-model classification pipelines tend to use. A pinned temperature: 0 plus a prefilled opening to force a label is precisely what now fails on two counts. The guide's replacement for format-forcing prefill is structured outputs or "tools with enum fields for classification," which is no help on Bedrock. Anyone budgeting from Haiku 4.5 token counts should recount with model set to claude-haiku-5-5 before trusting the headline price.

Primary sources: Claude Platform release notes, Claude Haiku 5.5 overview, What's new in Claude Haiku 5.5, Claude Haiku 5.5 migration guide, read 2026-10-13.