API Error 400: thinking blocks order error on first tool use after clean install

Same here

+1 on this. getting the same thing and it started within the last hour

+1

+1

Found 2 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/20691
2. https://github.com/anthropics/claude-code/issues/20689

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

to be honest, this is a much better and more detaield issue then mine. https://github.com/anthropics/claude-code/issues/20689

I can close mine and we can continue to pursue it here if that works.

Same issue:

Platform: darwin
Terminal: Apple_Terminal
Version: 2.1.19
Feedback ID: 74a044d8-6cbf-4c5b-881a-3ae9836f777a

Same issue

Also experiencing the same issue via Claude Code for VS Code

<img width="913" height="177" alt="Image" src="https://github.com/user-attachments/assets/dd705027-d063-4cf9-bf54-1b4f05c40277" />

Related?

Same issue here. Claude Code v2.1.19, macOS, Claude Max (Opus 4.5). Error occurs after tool use completes (e.g., 27 tool uses, 87.5k tokens). Fresh install and cache clearing didn't help.

same issue. started today. CC v2.1.19 macOS, claude max opus. tried /clear and restart CLI. stop at my first prompt which is asking it to read a md file

Read(setup.md)
⎿  Read 20 lines
⎿ API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0: If an
assistant message contains any thinking blocks, the first block must be thinking or redacted_thinking. Found
text."},"request_id":"req_011CXT8mjjFw7VwsDcDyTQ71"}

I'm getting the same. Just started about 15 minutes ago.

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.17.content.0.type: Expected `thinking` or `redacted_thinking`, but found `text`. When
     `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and `tool_result` blocks). We recommend you include
     thinking blocks from previous turns. To avoid this requirement, disable `thinking`. Please consult our documentation at
     https://docs.claude.com/en/docs/build-with-claude/extended-thinking"},"request_id":"req_011CXTCprBFh3zECAtZHyC4f"}

Claude says:

Looks like there's a persistent API issue with thinking blocks in the conversation history. That's not something I can fix from my side - it's related to how the client is
  constructing the API messages.

This is happening consistently with Opus 4.5 thinking on v2.1.19 on Linux.

same issue.

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0: If an assistant message contains any thinking
     blocks, the first block must be `thinking` or `redacted_thinking`. Found `text`."},"request_id":"req_011CXT7C2gffSy6TJrsgWVeV"} 

MacOS Sequoia 15.6.1
Claude code in CLI 2.1.19

Disabling "Thinking" in /config seems to help with the error though you loose the "benefits" of Thinking.

This is also happening to me without fail.

Downgrading from 2.1.19 to 2.1.15 did not fix the issue, suggesting it is not a CLI issue.

Disabling "Thinking" in /config seems to help with the error though you loose the "benefits" of Thinking.

Can confirm in my case that disabling thinking in the same chat without clearing context stopped the error.
MacOS Sequoia 15.6.1
Claude code in CLI 2.1.19

Yep, disabling thinking in /config "fixed" it, but you can feel the difference.

_More concerning than the issue itself, is the realization of how weird it is of not having Claude around, lol._

Yep, disabling thinking in /config "fixed" it, but you can feel the difference. _More concerning than the issue itself, is the realization of how weird it is of not having Claude around, lol._

I agree. I am also unsure how the claude cli got to a point where it does not implement the claude api correctly... even with plugins disabled I had the error.

Has been happening to me all afternoon. Also tried a fresh install, etc, but no luck

Same here. Downgrading doesn't resolve the issue.
The only thing that works is disabling thinking in the /config but not really a solution.

is it possible this is some backend side issue?
---
Environment:

• Claude Code version: 2.1.19 (also reproduced on 2.1.18)
• OS: Ubuntu 24 (6.8.0-90-generic #91-Ubuntu SMP PREEMPT_DYNAMIC Tue Nov 18 14:14:30 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux)
• Install method: native
• Account type: MAX

Error message:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0: If an assistant message contains any thinking blocks, the first block must be `thinking` or `redacted_thinking`. Found `text`."}}

Reproduction steps:

1. Start claude in any project (fresh or existing)
2. Complete 1-2 simple tasks successfully
3. Attempt a more complex task (e.g., creating files, multi-step work)
4. Error appears, persists for rest of session

Troubleshooting attempted:

• Cleared project-level .claude/ folder
• Cleared global ~/.claude/ and ~/.claude.json
• Re-authenticated
• Reinstalled Claude Code
• Tested on fresh /tmp project
• Downgraded to 2.1.18

Notes:

• Specific to Opus 4.5 (Sonnet works)
• Occurs consistently after extended thinking engages

I am surprised the status.claude.ai is not showing errors.

Now the error message suggest to disable thinking, so probably they are already aware and working on it
API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.17.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block (preceeding the lastmost set of tool_use and tool_result blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at
https://docs.claude.com/en/docs/build-with-claude/extended-thinking"}

Now the error message suggest to disable thinking, so probably they are already aware and working on it API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.17.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block (preceeding the lastmost set of tool_use and tool_result blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at https://docs.claude.com/en/docs/build-with-claude/extended-thinking"}

Probably Claude, its a Saturday.

Anyway, if you don't want to disable thinking, just keep prompting "Continue" the error doesn't seem to be systematic, so it can do a part of work, and then interrupt again. It is annoying but if you need to work with thinking activated this is the only solution at the moment.

Anyway, if you don't want to disable thinking, just keep prompting "Continue" the error doesn't seem to be systematic, so it can do a part of work, and then interrupt again. It is annoying but if you need to work with thinking activated this is the only solution at the moment.

Yea, I did this for a bit and burned through usage very quickly because a lot of work has to be "redone" just for it to fail again.

Same issue here. This is clearly a regression that started today (Jan 25) — multiple issues (#20691, #20698, #20711, #20717, #20725, #20726) all reported within hours. The thinking block ordering bug affects everyone using thinking mode with tool calls. Please upvote and comment to help get this prioritized!

same here

Another workaround is to use agents, it seems they are not affected by the bug.. It seems only the main claude get the error after using tools.

Seeing this error as well. Version 2.1.19

It seems the bug has been solved.

what is the "oncall" tag for?

Update: The issue appears to be fixed now! 🎉

Just tested extensively with thinking mode enabled (Opus 4.5) — 20+ consecutive tool calls including parallel operations, Task agents, WebSearch, file reads, Bash commands, etc. All succeeded without any 400 errors.

This likely relates to the Jan 22-23 incident on status.claude.com: "Elevated errors: Invalid signature in thinking block" which was marked resolved.

Can others please test and confirm if it's working for you now?

Update: The issue appears to be fixed now! 🎉 Just tested extensively with thinking mode enabled (Opus 4.5) — 20+ consecutive tool calls including parallel operations, Task agents, WebSearch, file reads, Bash commands, etc. All succeeded without any 400 errors. This likely relates to the Jan 22-23 incident on status.claude.com: "Elevated errors: Invalid signature in thinking block" which was marked resolved. Can others please test and confirm if it's working for you now?

Testing so far has been error free.

Can also confirm, seems things are working normally again.

I am no longer having the issue -- was able to visit the same chats that consistently returned errors and could "continue" them with thinking mode on with no error message.

Confirmed. It started 16 hours ago.

Version 2.1.19
Ubuntu 24
Installed w wget (not NPM)
Model: opus

I downgraded to 2.1.17 but didn't help
I removed all configs and tried in a new folder. Didn't help.

Using claude with --permission-mode=bypassPermissions -p

Got this very frequently today.

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.7.content.24: Invalid signature in thinking claude code

MacOS Sequoia 15.7.3.
VS Code extension: v2.1.22

I am getting this often today:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.11.content.21: thinking or redacted_thinking blocks in
the latest assistant message cannot be modified. These blocks must remain as they were in the original
response."},"request_id":"req_011CXjtBS9aLM6ZHhKFVbYig"}

When reading files or using subagents...

I am getting this often today: API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.11.content.21: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response."},"request_id":"req_011CXjtBS9aLM6ZHhKFVbYig"} When reading files or using subagents...

Looks like when using subagents.

We are also getting this today:

API Error: 400 {"detail":{"error":"{"message":"messages.1.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block (preceeding the lastmost set of tool_use and tool_result blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking. Please consult our documentation at https://platform.claude.com/docs/en/build-with-claude/extended-thinking

• Claude Code VS Code extension: v2.1.87
• LLM: AWS Bedrock Claude Sonnet 4
• OS: Ubuntu 24.04.4 LTS

I get the same error:

API Error: 400 messages.1.content.0.type: Expected `thinking` or `redacted_thinking`, but found `text`. When `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and
     `tool_result` blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable `thinking`. Please consult our documentation at https://docs.claude.com/en/docs/build-with-claude/extended-thinking

Claude Code Native 2.1.110

2+2 works fine though, seems like only thinking doesn't work:

<img width="3456" height="434" alt="Image" src="https://github.com/user-attachments/assets/334d1241-2753-466d-afee-cb1e0bf48810" />

The workaround that worked for me was to downgrade version to 2.1.86:
curl -fsSL https://claude.ai/install.sh | bash -s 2.1.86

I have the exact same problem today (Apr 23).

• Claude Code: 2.1.118, thinking mode enabled.
• Ubuntu 24.04.4 LTS.
• Problem occurs regardless of whether CC is used in VS Code or directly in bash.

----

I was tinkering with Bedrock configuration for an hour, no effect.

Downgrade to 2.1.86 as suggested by @emaysyuk is a viable workaround, problem goes away.
Edit: Even after downgrade, problem still persists when Claude Code is used from VS Code GUI.