claude-opus-5-5 returns 400 when disabled is passed; claude-sonnet-5-5 needs to switch to between_tools

Author’s note: this article is based on the Anthropic official docs dated 2026-11-08 (UTC+8) and the APIYI Claude thinking parameters guide updated at 10:40 on 2026-11-09, with all details fact-checked as of 2026-11-13. It’ll help you sidestep the 400 errors that claude-opus-5-5 and claude-sonnet-5-5 throw when you get the thinking parameter wrong.

If you pass disabled as the type for thinking on claude-opus-5-5, the API comes right back with a 400 — that’s the rule spelled out in the Anthropic official docs as of 2026-11-08 (UTC+8). APIYI (apiyi.com) refreshed its Claude thinking parameters guide at 10:40 on 2026-11-09, laying out the correct syntax for both model IDs. Here’s a quick rundown of what this change means for anyone calling the Claude API.

This isn’t rumor-mill stuff: both key conclusions come straight from the docs — claude-opus-5-5 can’t turn thinking off at all, you can only dial down effort; claude-sonnet-5-5 needs between_tools instead if you want to turn off pre-tool thinking.

The takeaway: 3 minutes gets you the correct syntax for both model IDs, the parameter boundaries of between_tools, and a troubleshooting path for that 400 error.

Core Info at a Glance: the thinking Rules for claude-opus-5-5 and claude-sonnet-5-5

Let’s start with the summary table, then go through it point by point:

Item	Details
Official source	Anthropic official docs, 2026-11-08 (UTC+8)
APIYI guide update	2026-11-09 10:40, already synced to docs.apiyi.com
claude-opus-5-5	Omit thinking and it uses adaptive thinking; pass disabled and you get a 400
claude-sonnet-5-5	Omitting thinking also means adaptive thinking; to turn off upfront thinking, use between_tools
effort default	medium for Opus 5.5, high for Sonnet 5.5
Max output	128000 for both model IDs, with thinking tokens counted inside that budget

The rules for claude-opus-5-5 are pretty straightforward. When a request doesn’t include the thinking field, the model goes into adaptive thinking and decides for itself whether to think this time and how deeply. But the moment you set thinking‘s type to disabled, the API doesn’t quietly ignore it — it returns a 400 outright. In other words, this model ID has no switch for turning thinking off. The only dial the docs give you is lowering effort, which defaults to medium.

claude-sonnet-5-5 is in a slightly different spot. It also supports adaptive thinking when thinking is omitted, but it has one extra officially blessed off-ramp: set type to between_tools, and you turn off upfront thinking while keeping only progress updates between tool calls. The boundaries of that path are spelled out just as clearly, so it’s worth checking the table below directly:

Constraint	What the docs say
effort values	Only accepts low, medium, high
Level ceiling	Can’t be higher than high
Combination limits	Can’t be sent together with display or budget_tokens
Actual effect	Turns off upfront thinking; keeps only progress updates between tool calls

Here’s a detail that’s easy to overlook: thinking tokens count toward max_tokens. Both model IDs have a max output of 128000, and if a request squeezes max_tokens too low, thinking can gobble up most of the budget first, and the actual response gets truncated. When you’re debugging a “why did the reply suddenly get so short” problem, it’s worth checking how much thinking is eating up first.

Figure note: the call chain in the diagram uses APIYI’s Anthropic-native-format gateway at apiyi.com as an example. When going through the Anthropic format, the base URL is https://api.apiyi.com — be careful not to write it in the /v1-suffixed form.

Impact Analysis: Who Gets Hit When disabled Returns a 400

This change in the wording hits callers in three places — error handling, parameter combinations, and budget planning. Breaking each one down makes it easier to see.

Impact on Existing Code: Migrating Away from disabled

Scenario	Old approach	What happens now	Correct approach
claude-opus-5-5 wants thinking off	Pass disabled as type	Returns 400	Can’t turn it off; only lower effort
claude-sonnet-5-5 wants interleaved thinking off	Pass disabled as type	Returns 400	Switch to between_tools
Wants fine-grained control over thinking budget	between_tools paired with budget_tokens	Combination isn’t accepted	The two can’t be sent together; use the three effort levels
Worried thinking will crowd out output	Setting max_tokens very low	Body text gets truncated	Thinking counts toward max_tokens, so leave headroom

The first impact is that parameter combinations have gotten more restrictive. between_tools can’t appear alongside budget_tokens or display, which means fine-grained thinking-budget control simply doesn’t work on this path — you’re left with indirect tuning through the low, medium, and high levels. Stack that on top of “thinking tokens count toward max_tokens” and you need to rebalance effort against max_tokens.

The second impact is that branching logic needs updating. If your script decides whether to send disabled based on model ID, then these two IDs need their own branches: the first one accepts the reality of “can’t turn it off” and steps effort down; the second switches over to between_tools. Keep that kind of branching in a single config spot so future wording changes touch as little code as possible.

One more thing worth noting: the same page says passing disabled to Fable 5.1 and Fable 5 also returns a 400 — they can’t be turned off either. Haiku 5.5 and Opus 5, meanwhile, can still use disabled as long as effort isn’t higher than high. If your code serves multiple model IDs at once, maintain a per-model toggle matrix instead of applying the same pattern everywhere.

Impact on Cost Structure: Thinking Tokens and Usage

On the APIYI (apiyi.com) master price table, claude-sonnet-5-5 is priced at $2 input, $10 output, and $0.1 cache read. Since thinking tokens count toward max_tokens, whether you turn off interleaved thinking and which effort level you pick both show up directly in token usage. In high-frequency call scenarios that gap gets amplified, so it’s worth running the numbers while you’re in there editing the code.

Caption: the wording was compiled from the official docs dated 2026-11-08; you can verify it against the original in the APIYI apiyi.com documentation center.

Model ID	disabled availability	Documented wording
claude-opus-5-5	Not available, returns 400	Thinking can’t be turned off; only lower effort (default medium)
claude-sonnet-5-5	Not available, returns 400	Use between_tools instead (default high)
Fable 5.1 / Fable 5	Not available, returns 400	Can’t be turned off either
Haiku 5.5 / Opus 5	Available when effort isn’t above high	The old syntax only holds for these two IDs

How to Quickly Verify between_tools on claude-sonnet-5-5

Before rolling out the change, it’s a good idea to validate the behavior difference with a small amount of traffic first. Register at APIYI (apiyi.com) and get an API key—new users get free test credits. Send requests to https://api.apiyi.com, set the type of thinking to between_tools in the request body, pick any one of the three effort levels, run one round, then compare against adaptive mode (where you don’t pass thinking at all), and watch how the output structure and token usage change.

🚀 Quick verification tip: Get free test credits through APIYI at apiyi.com, run the same set of prompts through both between_tools and adaptive modes, and confirm the difference before changing existing code. The platform supports unified invocation of mainstream models like Claude, OpenAI, and Gemini. The OpenAI-compatible interface is billed by usage, making it easy to compare and switch between models.

FAQ: disabled, between_tools, and 400 Errors

Q1: Can claude-opus-5-5 actually turn off thinking?

No. The official statement dated 2026-11-08 states clearly: passing `disabled` returns a 400, and this model ID has no switch to turn off thinking—you can only reduce thinking depth via `effort`, whose default level is `medium`. If your scenario absolutely requires no thinking at all, you’ll have to switch model IDs.

Q2: Can between_tools on claude-sonnet-5-5 be combined with budget_tokens?

No. The docs explicitly state that `between_tools` can’t carry `display` or `budget_tokens` at the same time, and its corresponding `effort` only accepts three levels—`low`, `medium`, and `high`—and can’t exceed `high`. If you want fine-grained control over the thinking budget, you’ll have to fall back to adaptive mode or switch model IDs.

Q3: Why did my Claude request suddenly return a 400?

First check whether you set the `type` of `thinking` to `disabled` in the request body—on claude-opus-5-5 and claude-sonnet-5-5, that’s the documented trigger condition. Once confirmed, handle it per the comparison table in this article: for the former, lower `effort`; for the latter, switch to `between_tools`. The applicable syntax is based on the official documentation dated 2026-11-08 (UTC+8), and the APIYI page was synced on 2026-11-09 at 10:40.

Q4: How do I quickly verify the effect of between_tools on claude-sonnet-5-5?

Register an account at APIYI (apiyi.com) and get an API key—new users get free test credits. Send requests to `https://api.apiyi.com`, set the `type` of `thinking` to `between_tools`, pick any one of the `effort` levels, and run one round to compare the difference against adaptive mode.

Q5: Do thinking tokens eat into the output budget?

Yes. Thinking tokens count toward `max_tokens`, and the maximum output for both model IDs is 128000. If you set `max_tokens` too low in a request, thinking may consume the budget first and the main content gets truncated, so it’s best to reserve enough room for the main content.

claude-opus-5-5 and claude-sonnet-5-5: Key Takeaways

• claude-opus-5-5 won’t let you turn off thinking: passing disabled gets you an immediate 400 — the only option is to dial it down with effort, which defaults to medium.
• claude-sonnet-5-5 switched to between_tools: this turns off upfront thinking and keeps only the progress updates that happen between tool calls.
• Three parameter red lines: effort accepts only low, medium, and high — nothing above high — and it can’t be combined with display or budget_tokens.
• Budget rule: thinking tokens count toward max_tokens, and both IDs cap out at 128000.
• Where to verify: APIYI apiyi.com. The Anthropic-format base is https://api.apiyi.com — don’t add /v1.

Summary: three moves that keep you out of 400 territory

The whole story on claude-opus-5-5 and claude-sonnet-5-5 boils down to three sentences:

1. claude-opus-5-5: passing disabled triggers a 400, so thinking can’t be turned off — you can only lower effort.
2. claude-sonnet-5-5: turn off upfront thinking by switching to between_tools, stick to the three effort levels, and don’t pair them with display or budget_tokens.
3. Budget: thinking tokens count toward max_tokens, which tops out at 128000 — so leave some headroom for the actual response.

If you’re stuck on this 400 right now, the quickest path is to head to APIYI apiyi.com, grab the free test quota, and run a round of between_tools to confirm the behavior before you touch your existing code. One API key on the platform gets you OpenAI, Claude, Gemini, and other mainstream models, billed by usage — handy for quick comparisons and switching.

📚 References

1. APIYI Claude Thinking Parameters Guide: How to use effort and thinking
2. Link: docs.apiyi.com/api-capabilities/claude-effort-thinking.md

3. Description: The version updated on 2026-11-09 10:40; the main source of facts for this article

4. APIYI Guide Refresh Update: Official notes on this round of alignment

5. Link: docs.apiyi.com/live/2026-11/claude-thinking-guide-refresh.md

6. Description: Documents the process of aligning with the official documentation on 2026-11-08

7. Anthropic Official Documentation: The original wording for the thinking parameter

8. Link: docs.anthropic.com
9. Description: Recommended for cross-reading to confirm the latest usage

---

Author: APIYI Technical Team
Technical discussion: Feel free to reach out via the APIYI website homepage. For more materials, visit the docs.apiyi.com documentation center.

Related Articles

• Claude Fable 5.2 vs. Claude Opus 5.5: How to Choose: Claude Fable 5.2 vs. Claude Opus 5.5: At twice the standard price, it doesn’t win in every category; this article offers a task-based model selection table covering specs, pricing, default effort level, and release benchmark results.
• Sonnet 4.5 will be retired on November 30: A 5-Step Guide to Migrating from claude-sonnet-4-5 to Sonnet 5.5: Claude Sonnet 4.5 will be retired on 2026-11-30, and this article uses a comparison table and a five-step migration sequence to explain changes to the model ID, cache read multiplier, and billing when migrating to Claude Sonnet 5.5, as well as how to verify the migration with the same request afterward.