Porting a Sume agent loop to Haiku 5.5: drop temperature, top_p, top_k
Anthropic says to omit temperature, top_p and top_k on Haiku 5.5, and some values return a 400. Strip them from your loop before the first Sume tool call.

Remove temperature, top_p and top_k from every request that targets claude-haiku-5-5. The Haiku 5.5 overview tells you to omit them and points to a migration section listing the values that return a 400. If your tool-calling loop for Sume was written for an older model with temperature=0 set for repeatability, the first request after the model swap can fail for that reason alone.
Port checklist
These are the items on the model page that change a loop. None of them concerns Sume's API.
| Item | Haiku 5.5 |
|---|---|
| Model id | claude-haiku-5-5 |
| Sampling parameters | Omit temperature, top_p, top_k |
| Thinking | Adaptive, on by default; steer with effort (default medium) |
| Tokenizer | Same text is about 30% more tokens than on Haiku 4.5 |
| Context / output | 1M tokens / 128K |
What repeatability means now
You no longer get determinism from temperature, so make the Sume side safe to repeat instead. Every paid call carries an idempotency_key derived from your own order id; replaying the same turn then returns the original job and does not create a second one. See the tools and gates page.
Budget for the tokenizer: a tool list that was a given size on Haiku 4.5 is about 30% larger in tokens on 5.5, which matters near the 100,000-token price step.
A one-hour port
Search your codebase for the three parameter names and for any wrapper that injects them. Remove them for the Haiku 5.5 branch only; leave them in place for older models while both are running. Change the model id to claude-haiku-5-5, run one tool call against Sume's read-only tools such as balance_get and check the response before enabling any paid tool.
Then check effort. Adaptive thinking is on by default at medium; a polling agent does not need more, and a lower setting is cheaper. Test that your loop's stop conditions still work.
- Read-only smoke test first.
- Branch by model id.
- Re-measure tokens after the swap.
Sources
Related posts
More in Models
- HelixBench: what spatial-acoustic consistency tests in AI video
HelixBench scores whether synthesized sound fields track camera movement. What the arXiv page says it measures, and a manual way to spot-check any AI clip.
- HelixWorld license: Apache 2.0 code, weights and paper differ
HelixWorld's code is Apache 2.0, its weights have a separate license on Hugging Face, and the arXiv page lists CC BY-NC-SA 4.0. Read all three.
- Image-to-video API with sound: five-second cost by Sume row
Omni Flash 1.1 (always on), Wan 3.0 (audio), MiniMax H3 and H3 Max (native stereo) take a first frame and render sound. A 5 s clip is $0.19 to $0.94 on Sume.
- Is Qwen Image 2.1 Turbo on Sume? No: use these three Qwen rows
Sume's image catalog has no Qwen-Image-2.1 Turbo row. It lists Qwen Image, Qwen Image Max and Qwen Image Layered, and one call shows which.
Written by Sume