[BUG] Max 5x: Session limits depleting 3-4x faster than normal since June 24, 2026 (claude.ai web)

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/69892
2. https://github.com/anthropics/claude-code/issues/69904
3. https://github.com/anthropics/claude-code/issues/69430

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

same here 100% of my 5-hour max plan limit used on claude opus 4.6 max that would normally have even been 10% of my total usage for the time used. Not to mention a significant drop in the quality of responses, more behind the scenes tampering without communication, not cool anthropic!

This isn't stale. It's been an issue and continues to have issues opened.

A measured data point for Max 5x, instead of impressions, from 18 September 2026 (Claude Code on Opus 5). It's only a lower bound, but it puts a number on what one percent of the 5-hour window buys.

Method

• The desktop app samples the usage bars locally every ~15 minutes in ~/Library/Application Support/Claude/plan-usage-history.json: fh is the 5-hour window and sd the 7-day window, both in percent.
• Claude Code writes token usage for every model response to ~/.claude/projects/*/*.jsonl. I deduplicated by message id, because each response is written several times (see #91775), and priced it at Opus 5 rates of $5 input / $25 output / $10 cache write / $0.50 cache read per million tokens. Those rates fit the cost records of 113 of my desktop sessions exactly.
• For each sampling interval in which Claude Code spent anything and the window didn't reset, I compared the dollars spent with the rise in fh.

Result: 7 intervals, $42.11 of API-priced usage, fh up 57 points.

| Interval | API-priced spend | 5-hour % gained | $ per 1% |
|---|---:|---:|---:|
| 15:05–15:20 | $3.21 | +5 | 0.64 |
| 15:20–15:35 | $7.69 | +6 | 1.28 |
| 16:50–17:05 | $4.78 | +4 | 1.20 |
| 17:05–17:20 | $2.82 | +3 | 0.94 |
| 19:05–19:20 | $9.68 | +15 | 0.65 |
| 19:20–19:35 | $6.12 | +12 | 0.51 |
| 23:05–23:20 | $7.81 | +12 | 0.65 |

Pooled, that is $0.74 per 1% (median $0.65), so at least ~$65–75 of API-priced Opus usage per full 5-hour window on Max 5x on that day. The 7-day bar moved only 6 points over the same intervals, so a weekly figure (~$700) is very rough.

Why it's a lower bound: most of that day's window consumption doesn't appear in local logs. fh rose another 273 points across 39 intervals with no Claude Code activity at all, which would be claude.ai chat, desktop agent sessions or other devices. If any of that overlapped the 7 intervals above, it inflated the percentage without adding visible dollars, so the true value per 1% is at least this high.

How to measure it cleanly (if anyone wants to): at the start of a fresh 5-hour window, with nothing else running, make a few claude -p ... --output-format json calls of known total_cost_usd, then read fh from that file after the next sample. Repeated across days, that would show directly whether the dollars per percent change over time, which is the question this issue is about.