[BUG] 5-Hour Rate Limit Exhaustion Accelerating Despite Comparable Workload

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?

The 5-hour rate limit is being exhausted 3x faster today compared to previous days with comparable workloads. This represents a service degradation that violates expected usage patterns and contractual rate limit behavior.
Evidence of Acceleration
Date Total Requests Input Tokens 5h Exhaustion Events Time to First Exhaustion
2026-03-21 ~900 ~3M 0 N/A
2026-03-22 ~900 ~3M 0 N/A
2026-03-23 908 3,402,759 0 N/A
2026-03-24 936 (+3.1%) 4,169,220 (+22.5%) 3 ~4.5 hours

Key Finding: A 3-23% workload increase should not cause infinite → 4.5 hour exhaustion time.
Rate Limit Exhaustion Log

2026/03/24 13:39:46 5h=100.0% 7d=15.0% → 429 Too Many Requests
2026/03/24 16:02:48 5h=100.0% 7d=24.0% → 429 Too Many Requests
2026/03/24 16:34:37 5h=100.0% 7d=24.0% → 429 Too Many Requests

5h Window Progression (2026-03-24)
Time (UTC) 5h% Time Since Reset
09:24:19 1.0% —
10:09:42 11.0% —
13:39:46 100.0% 4h 15m
15:11:09 57.0% (reset) —
16:02:48 100.0% 51m

The 5h window reset and re-exhausted in 51 minutes.
Cache Efficiency Degradation
Date Cache Creation Tokens vs Input Ratio
2026-03-23 5,698,495 1.68x
2026-03-24 7,287,796 1.75x
Same workload burning 28% more cache creation tokens, accelerating rate limit exhaustion.
Breach of Service Agreement
Issue: Unilateral Service Degradation

The acceleration of rate limit exhaustion represents a material change to service behavior without:

Advance notice to users
Documentation updates
Corresponding workload increase to justify the change

Quantified Impact
Metric 2026-03-23 2026-03-24 Delta
5h exhaustions 0 3 ∞ increase
Hours of service ~24 ~7 (cumulative) -71%
Cache creation tokens 5.7M 7.3M +28%
Cost efficiency Baseline Degraded Unknown excess cost
Request for Recompensation

We request the following:

Credit for service downtime: 3 complete 5h window exhaustions = ~15 hours of unavailable service capacity

Refund for excess token burn: 7.3M vs expected ~6M cache creation tokens = ~1.3M excess tokens at published rates

Investigation commitment: Acknowledgment of this issue and timeline for fix

Transparency: Explanation of what changed between 2026-03-23 and 2026-03-24 to cause this acceleration

What Should Happen?

Expected Behavior (Contractual)

Consistent Rate Limit Behavior: The 5h rate limit should exhibit consistent exhaustion timing for comparable workloads, absent documented service changes.

Documented Thresholds: Rate limit behavior should match published documentation. Acceleration of ~28% in token burn rate without corresponding workload increase suggests a service-side issue.

Graceful Degradation: If rate limits change, users should receive advance notice or documentation updates.

Actual Behavior (Breach)
Aspect Expected Actual
Exhaustion timing ~12-24h for this workload 4.5h, then 51m
Consistency day-over-day Similar exhaustion pattern 0 → 3 exhaustions
Cache efficiency Stable ratio to input tokens +28% increase

Error Messages/Logs

2026/03/24 13:39:46. [UPSTREAM] /v1/messages 429 Too Many Requests
{"type":"error","error":{"type":"rate_limit_error","message":"This request would exceed your account's rate limit. Please try again later."},"request_id":"req_"}


2026/03/24 16:02:48. [UPSTREAM] /v1/messages 429 Too Many Requests
{"type":"error","error":{"type":"rate_limit_error","message":"This request would exceed your account's rate limit. Please try again later."},"request_id":"req_"}


2026/03/24 16:34:37. [UPSTREAM] /v1/messages 429 Too Many Requests
{"type":"error","error":{"type":"rate_limit_error","message":"This request would exceed your account's rate limit. Please try again later."},"request_id":"req_"}

Steps to Reproduce

Run Claude Code with standard workload (~900 requests, ~3-4M input tokens per day)
Observe 5h rate limit progression via response headers or proxy logging
Compare day-over-day exhaustion timing

Baseline Day (2026-03-23):

908 requests
3.4M input tokens
5.7M cache creation tokens
Result: 0 exhaustions

Degraded Day (2026-03-24):

936 requests (+3%)
4.2M input tokens (+22%)
7.3M cache creation tokens (+28%)
Result: 3 exhaustions in 7 hours

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

666

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Xterm

Additional Information

This is systemic and currently in progress. We the users demand compensation.