[BUG] Team Premium seat weekly limit stuck at 100% after only 102M tokens — other users on same plan consume 757M+ without issues

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?

A Team Premium seat ($150/month) provisioned on Friday March 27, 2026 shows 100% weekly limit consumed by Saturday March 29, while session usage shows 0%. The user is completely blocked.

Important context: The user previously had a personal Pro plan. The Team Premium seat was enabled 3 days ago. Cached credentials from the old Pro plan may be contributing to the rate limiter associating this account with incorrect limits.

We monitor all Claude Code traffic through a LiteLLM proxy and have exact token counts:

| Model | Successful Requests | Failed | Tokens |
|-------|-------------------|--------|--------|
| anthropic/claude-opus-4-6 | 720 | 0 | 96,412,699 |
| anthropic/claude-haiku-4-5-20251001 | 151 | 0 | 6,230,830 |
| cc-team-claude-opus-4-6 | 0 | 13 | 0 |
| Total | 871 | 13 | 102,643,529 |

For comparison, another user on the same Team Premium plan consumed 757,835,578 tokens (7.4x more) with 6x more requests and has not reached 50% of their weekly limit.

The 13 failed requests on cc-team-claude-opus-4-6 are the rate limit errors themselves.

Usage peaked on March 29 (~70M tokens, ~500 requests) and dropped to zero on March 30 when the block took effect. The claude.ai dashboard currently shows:

• Session: 0% used
• Weekly limit (all models): 100% used

This matches the server-side rate limiter bug reported in #38335, #39649, and #39581.

What Should Happen?

Weekly limit should reflect actual token consumption. A user consuming 102M tokens should not be blocked while another user consuming 757M tokens on the identical plan is unaffected.

Error Messages/Logs

Claude AI usage limit reached, please try again after [reset time]

claude.ai Settings > Usage:
- Session: 0% used
- Weekly limit (all models): 100% used

LiteLLM proxy logs show 13 failed requests on cc-team-claude-opus-4-6 (rate limit errors)
All 871 successful requests completed normally before the block.

Steps to Reproduce

1. Provision a new Team Premium seat ($150/month) on March 27, 2026
2. User authenticates Claude Code via OAuth ("Claude account with subscription"), routed through LiteLLM proxy for monitoring
3. User works normally over March 27-29: 720 Opus requests + 151 Haiku requests = 871 total successful requests, 102M tokens
4. On March 30, weekly limit shows 100% consumed on claude.ai Settings > Usage
5. Session shows 0% — no active session consuming tokens
6. User is completely blocked from using Claude Code
7. Compare with another user on the same Team Premium plan in the same organization: 757M tokens (7.4x more), 6x more requests, not even at 50% weekly usage

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

2.1.52 (when issue started)

Claude Code Version

2.1.87 (updated after issue, problem persists)

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

VS Code integrated terminal

Additional Information

• User previously had a personal Pro plan before being provisioned on the Team Premium seat 3 days ago. The rate limiter may be applying the old Pro plan limits instead of the Team Premium limits. This matches bug #40340 where cached credentials from a previous plan persist until re-login.
• We have exact LiteLLM proxy logs proving the token counts. Screenshots with token breakdown by model and daily request distribution are available.
• Organization is on a Team plan with multiple Premium seats — only this specific account is affected. All other users on the same plan work normally.
• The only Anthropic-related env var set is ANTHROPIC_CUSTOM_HEADERS (for LiteLLM routing). No ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, or CLAUDE_CODE_OAUTH_TOKEN — OAuth credentials are not being overridden.
• Claude Code was v2.1.52 when the issue started. Updated to v2.1.87 — problem persists.
• Support ticket already submitted via claude.ai (Conversation ID: 215473699851042), escalated to human specialist.
• Related issues: #38335, #39649, #39581, #29579, #16157, #40340