[BUG] Thinking tokens do not fire reliably or visibly in Claude Code web sessions

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Extended thinking appears to not activate consistently during Claude Code sessions, even when effort level is set to high. There is no visible indicator in the session when thinking tokens are being used, making it impossible for the user to verify whether reasoning is actually happening.

Thinking tokens appear to not fire, or fire inconsistently. Even when thinking may have occurred, the model cannot see or reference its own thinking content from previous messages - thinking blocks are not retained in the conversation context. This means there is no way for either the user or the model to verify after the fact whether reasoning happened, or to catch and correct flawed reasoning from a prior turn. The model defaults to pattern-matched responses rather than reasoned ones.

What Should Happen?

When extended thinking is enabled, thinking tokens should fire reliably on non-trivial tasks, and there should be some visibility or confirmation that they are being used.

Error Messages/Logs

Steps to Reproduce

1. Open claude.ai/code and start a session

1. Set effort level to high

1. Give the model a non-trivial multi-step task (e.g. a git workflow task with specific instructions)

1. Observe that no thinking block is visible before or during the response

1. After the response, ask the model to reproduce or reference its thinking content from that message

1. Observe that the model cannot - it either admits it didn't think, or fabricates a plausible-sounding but unverifiable description

1. Ask the model to confirm whether thinking tokens fired at all - it cannot confirm either way

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

-

Platform

Other

Operating System

Other

Terminal/Shell

Other

Additional Information

Workaround:

Explicitly instructing the model to write out its reasoning in [mock thinking]...[/mock thinking] blocks before responding. This forces visible, structured reasoning and noticeably improves response quality - but should not be necessary.

Context:

• Plan: Claude Max ($200/month)

• Effort level: set to high

• Environment: claude.ai/code (web)

• Model: claude-sonnet-4-6

Impact:

Users paying for premium tiers expect extended thinking to work as advertised. Without it, response quality degrades significantly on tasks requiring careful reasoning.