Claude AI usage limit prematurely reached

I’ve been using Claude Code and ran into an issue with session limits that doesn’t seem consistent with the documented token caps.
• Session start: 8:00 AM
• Elapsed: ~1 hour
• Tokens used: ~8.7M (about 31% of the stated 28M session limit)
• Burn rate: ~155k tokens/min (flagged as “normal”)
• Reset time: 1:00 PM

Despite being far below the stated token allocation (only ~31%), Claude Code blocked further usage with the message:
USAGE LIMIT. RESET AT 01:00 PM
PROJECTION: WILL EXCEED LIMIT (148%)

This feels misleading, because:

1. I had only been in the session for 1 hour.
2. My actual usage was well under the advertised cap.
3. The system seems to be cutting me off based on projected burn rate, not actual usage.

While I understand the need for rate limiting, the way this is surfaced is confusing. The usage data clearly shows that I should still have plenty of room before hitting the limit, but I’m forced to wait for the reset window anyway.

I notice that this usage limit message appears as soon as it automatically compacts my message. Is the compacting process itself consumes a lot of usage? If so, this is not fair as it is an automatic process.

Feature requests / questions:

1. Is the session cap supposed to be strictly time-based (5 hours) or token-based (28M), or a mix of both?
2. Would it be possible to allow users to continue until the actual limit is hit, rather than blocking preemptively?

<img width="1352" height="844" alt="Image" src="https://github.com/user-attachments/assets/ef8f7ceb-4b3d-4159-99aa-54e277984f17" />