Max Plan 5x: 143x token usage drop, limit reached with 12M tokens (was 1,714M without issue)

Bug Description

Max Plan 5x user experiencing dramatic rate limit reduction since ~March 28, 2026. Same type of work that previously consumed 1,700M+ tokens without hitting limits now triggers rate limits at just 12M tokens — a 143x decrease in effective capacity.

Environment

• Plan: Max Plan with 5x extra usage enabled
• Claude Code version: 2.1.92 (latest)
• OS: macOS Darwin 25.3.0 (Apple Silicon)
• Model: opus-4-6

Hard Data (from local JSONL analysis)

Daily token usage — before vs after:

| Date | Messages | Output Tokens | Cache Read | Total Tokens | Hit Limit? |
|------|----------|--------------|------------|-----------------|------------|
| 03-25 | 1,977 | 788K | 585M | 660M | No |
| 03-26 | 1,656 | 524K | 573M | 613M | No |
| 03-27 | 1,998 | 1,296K | 1,702M | 1,714M | No |
| 03-28 | 668 | 339K | 148M | 153M | No |
| 03-30 | 2,060 | 1,310K | 785M | 802M | No |
| 03-31 | 545 | 565K | 127M | 134M | Partial |
| 04-03 | 857 | 785K | 246M | 252M | Partial |
| 04-05 | 114 | 106K | 11M | 12M | Yes |

On April 5th, with 17x fewer messages, 12x less output, and 55x fewer total tokens — the rate limit was hit.

Session-level evidence:

The longest sessions from the first week show healthy cache efficiency:

• 03-25 (1,236 msgs, 6h): cache ratio 348x
• 03-26 (1,578 msgs, 9h): cache ratio 513x
• 03-27 (1,982 msgs, 16h): cache ratio 302x

Recent sessions show the same pattern at their scale (short sessions naturally have lower ratios), but the total consumption is dramatically lower while limits are hit faster.

What Changed

• No configuration changes that would explain this — skills were added then removed, MCP servers cleaned up, settings optimized. Problem persists regardless.
• Claude Code version kept up to date
• Same type of work — Python development, web research, code editing

Expected Behavior

With Max Plan 5x enabled, usage capacity should remain consistent. A day producing 12M total tokens should not hit rate limits when previous days at 1,714M tokens did not.

Related

• Likely related to #38335 (cache bug causing inflated token billing)
• Anthropic acknowledged on March 31 that "people are hitting usage limits in Claude Code way faster than expected"
• Thariq Shihipar announced intentional peak-hour throttling on March 26, but this affects ALL hours, not just peak

Request

Please investigate whether:

1. The rate limit formula changed for Max Plan 5x users around March 28-31
2. Cache read tokens are being counted at a higher weight than before
3. There are hidden token costs (e.g., thinking tokens) not reflected in the JSONL usage data but counted against rate limits

Data collected using ccusage and direct JSONL analysis from ~/.claude/projects/.