[BUG] Incorrect token limit exceeded error despite low usage

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?

We are experiencing an issue where Claude returns a "token limit exceeded" (or equivalent) error, even though the actual usage metrics indicate that we are well below the allowed limits.

Observed behavior:

• Claude stops processing requests and reports that all tokens have been consumed.
• However, the usage dashboard shows significantly lower consumption than the reported limit.

Evidence:
From the usage metrics:

• Total Requests: 577
• Total Successful Requests: 569
• Total Tokens: ~38,755,230
• Total Spend: ~$31.90

These numbers do not align with hitting any expected token limit.

Expected behavior:
The system should only return a token limit exceeded error when the actual usage reaches the configured quota.

Impact:

• Blocks normal usage of Claude
• Causes confusion due to mismatch between reported limits and actual metrics
• Potential disruption in production workflows

Additional context:
This appears to be either:

• A miscalculation of token usage
• A caching/state issue on the quota system
• Or an incorrect limit being applied to the account

Request:
Please investigate the discrepancy between reported token exhaustion and actual usage metrics, and clarify which limits are being enforced.

Attachments:
(Screenshot showing usage metrics vs error)

What Should Happen?

Claude should continue processing requests normally as long as the actual token usage remains below the configured quota limits.

The system should:

• Accurately track token consumption
• Only return a "token limit exceeded" error when the real usage reaches the defined limit
• Keep the API behavior consistent with the usage metrics shown in the dashboard

There should be no discrepancy between reported usage and enforced limits.

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce:

1. Send multiple requests to Claude via API (normal usage, no unusual payload sizes).
2. Continue sending requests until the system starts returning a "token limit exceeded" (or similar) error.
3. Open the Anthropic usage dashboard.
4. Compare the reported token usage and spend with the expected limits.
5. Observe that the error appears even though the usage metrics are significantly below any reasonable quota.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.87 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Warp

Additional Information

_No response_