[BUG] Research mode withdraws ~68% of session usage instantly on first message, even when request errors with no output (Pro plan)

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?

Bug Description

Enabling Research mode and sending it as the first message of a brand-new session — zero prior token usage, completely blank conversation — causes an immediate, disproportionate withdrawal of session usage before any research is performed. The request errors out and produces no output, yet the usage deduction still occurs. This happens instantly upon the server receiving the request, not as a result of processing time, conversation length, or context window size.

This is repeatable and consistent, not a one-off failure or edge case tied to a specific prompt.

Actual Behavior

The full usage deduction happens instantly upon the server receiving a Research-mode request, independent of whether the request completes, errors, or returns any output. This occurs on a completely fresh session with no other usage factors (message length, conversation length, prior context) in play, meaning it isn't explained by any of the officially documented usage-limit factors.

Severity: High, possibly fraud

This directly affects paid usage on the Pro plan. Session usage is a finite resource tied to a paid subscription. Withdrawing a large share of that usage (~68% in this case) for a request that failed and delivered nothing is a billing integrity issue, not a cosmetic bug, and should be treated as high priority given it is reproducible rather than an isolated incident. Whats worse is that this indiscriminately just swallows up credit usage, vanishing users' paid credits with no work performed. This is very obviously deceitful, programmed, undocumented behavior

Environment

• Plan: Pro
• Interface: claude.ai, claude mobile, claude.ai on desktop and on mobile
• Trigger condition: Research mode enabled on the first message of a new session

What Should Happen?

Usage should be deducted in proportion to tokens actually processed and/or output actually generated. A request that errors server-side and returns no output should not withdraw a large share of session usage.

Error Messages/Logs

If more than ~30% of session tokens are used, the research mode will fail immediately and produce a warning that "You've used your session tokens" or whatever it says, blocking any more messages until reloading the page

Steps to Reproduce

1. Start a brand-new conversation on claude.ai (Pro plan) with no prior messages in the session.
2. Enable Research mode.
3. Send a single prompt with Research mode enabled as the first message of the conversation.
4. Observe the request error out, producing no research output and no completed response.
5. Check the session usage indicator immediately after: a large percentage of session usage (~68% of the session, roughly 1.5–2 million tokens) has already been withdrawn.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

1.2.3

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

I have reached out to costumer support about this issue and obviously received no response despite being promised a response from "a real person" and instead only being sent an ironic satisfaction survey by email.