What breaks when you move code from Haiku 4.5 to claude-haiku-5-5, the fix for each, and what to re-tune.
Model ID: claude-haiku-5-5 (Amazon Bedrock: anthropic.claude-haiku-5-5)
Pricing: $0.10 / $0.50 per million input / output tokens for prompts up to
100K tokens. $0.50 / $2.50 above 100K tokens. (Haiku 4.5: $1 / $5 at any
length.)
Context / output: 1M tokens / 128K tokens (Haiku 4.5: 200K / 64K)
Last verified: 2026-10-11
Source: Anthropic API migration guide (bundled claude-api skill, cached
2026-10-06) and platform docs
This is the current Haiku, for high-volume and latency-sensitive work: classification, routing, extraction, and sub-agent tasks.
effort levels (low to max, default medium).| Change | Detail |
|---|---|
| Manual thinking budget | {type: "enabled", budget_tokens: N} 400s. Use {type: "adaptive"} (or omit) plus output_config.effort. |
| Sampling parameters | temperature must be 1, top_p must be 0.99, any top_k 400s, and sending both temperature and top_p 400s. Omit all three. A temperature=0 classifier: remove it and constrain output with structured outputs or an enum tool. |
| Assistant prefill | A final assistant turn 400s, even with thinking off. End messages with a user turn. Use output_config.format, an enum tool, or a prompt instruction. |
| Computer use | computer_20250124 400s on the Claude API and Google Cloud. Use computer_toolset_20260801. Replace the fine-grained-tool-streaming-2025-05-14 header with eager_input_streaming: true. Do not assume the Bedrock exception that Opus 5.5 and Sonnet 5.5 have. |
| Preserved thinking | Replaying a thinking block after editing system, tools, or earlier messages 400s on accounts created on or after 2026-08-31. Keep histories append-only. |
| Refusals | Four categories: cyber, bio, frontier_llm, general_harms. Check stop_reason first. Do not send fallbacks: "default" leaves a decline declined, and a model list 400s. Handle it in the client. |
| Change | Action |
|---|---|
| Thinking on by default | Read content blocks by type, never content[0]. A response can start with thinking blocks. |
Thinking counts toward max_tokens |
A cap sized for a one-word answer (for example 50) can be consumed by thinking. Raise it or lower effort. |
| Thinking text omitted | Set display: "summarized" to see reasoning. Keep empty thinking blocks. |
| Tokenizer | Recount with count_tokens against claude-haiku-5-5. Re-baseline max_tokens and cost. |
| Disabled thinking | Still accepted at low, medium, high. 400s at xhigh and max. Prefer adaptive at low. |
Forced tool_choice |
Still accepted, but the response skips thinking. Use auto plus a prompt when the call needs thought. |
| Cache minimum | 512 tokens (was 4,096). Short prompts now create cache entries. |
| Thinking blocks | Stay with the account that produced them. Replay a stored conversation through the same account. |
| Level | Use |
|---|---|
low |
Chat, short tool tasks, simple high-volume routes. Cheapest and fastest. |
medium (default) |
Most work, including agentic coding. |
high |
Knowledge work, long agent tasks, strict instruction following. |
xhigh, max |
Only with a measured gain. Compare against Sonnet 5.5 first. |
To get less thinking, lower effort. A “answer directly” prompt barely helps.
low effort, add a line saying which facts may have changed
since training. Avoid blanket “always search” rules.xhigh in multi-turn chat.low effort can stop early. Add a keep-working
instruction, or raise effort.low and medium.tool_result.| Version | Model ID | Status |
|---|---|---|
| Haiku 5.5 | claude-haiku-5-5 |
Current |
| Haiku 4.5 | claude-haiku-4-5 |
Active |