[BUG] Max 20x plan ($200/mo) — quota exhausted in ~1h, making the product unusable for our team

[TrendSys]

We are a development team with 1 Max 20x subscription ($200/mo) and 2 Max 5x subscriptions ($100/mo each) — totaling $400/month on Claude Code.

Since approximately March 23, 2026, our Max 20x plan quota has been exhausting in roughly 60–90 minutes with the exact same usage patterns that previously lasted the full 5-hour window without issues. This is happening consistently, not just during peak hours.

Impact

• Our team cannot work. We are blocked multiple times per day waiting for quota resets.
• Usage behavior has not changed — same types of tasks, same frequency, same codebase.
• We have tried switching between Opus and Sonnet, starting fresh sessions, and updating to the latest version. The problem persists.

What we've tried

• Updated Claude Code to latest version (v2.1.91, autocompact fix included)
• Alternating between Sonnet and Opus to spread across separate limits
• Avoiding peak hours (5am–11am PT)
• Shorter sessions with fresh conversations

Our concern

We are paying $400/month across 3 subscriptions and the product has become unusable. We are actively evaluating alternatives from other providers (Cursor, Windsurf, GitHub Copilot) and considering cancelling all 3 subscriptions if this is not addressed transparently.

What we need

1. Acknowledgment that this is a known issue affecting Max 20x subscribers
2. Transparency on root causes beyond the autocompact fix — the problem persists post-patch
3. A timeline for resolution or, at minimum, honest communication about capacity constraints
4. Consideration for affected subscribers — credit, extended limits, or some form of remediation for the weeks of degraded service

Environment

• Plan: Max 20x ($200/mo) + 2× Max 5x ($100/mo)
• Claude Code version: 2.1.91
• Model: Opus 4.6 (1M context)
• OS: Windows 11
• Related issues: [BUG] Claude Max plan session limits exhausted abnormally fast since March 23, 2026 (CLI usage) #38335, Max 20 plan: rate limit 100% exhausted within ~70 minutes after reset (never happened before v2.1.89) #41788, [BUG] Critical: Widespread abnormal usage limit drain across all paid tiers since March 23, 2026 — multiple root causes identified, no formal communication issued #41930, [Bug] Auto-Compact loops continuously in v2.1.89 #41767

We appreciate the product when it works — but right now, it doesn't. Please communicate with your paying customers.

[github-actions]

Found 3 possible duplicate issues:

1. Max 20 plan: rate limit 100% exhausted within ~70 minutes after reset (never happened before v2.1.89) #41788
2. [BUG] Claude Max plan session limits exhausted abnormally fast since March 23, 2026 (CLI usage) #38335
3. [BUG] Critical: Widespread abnormal usage limit drain across all paid tiers since March 23, 2026 — multiple root causes identified, no formal communication issued #41930

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

[TrendSys]

This is not a duplicate. The referenced issues report the same symptoms, but none of them represent a team paying $400/month across 3 subscriptions (1× Max 20x + 2× Max 5x) that is about to churn.

The existing issues (#38335, #41788, #41930) are already weeks old with no official Anthropic response. We deliberately reference them to show the problem is widespread and unresolved — not to duplicate them.

Our issue adds:

• Business impact: 3 paid subscriptions at risk of cancellation
• Post-patch confirmation: the problem persists on v2.1.91 (after the autocompact fix)
• Team perspective: multiple developers affected simultaneously, not just a single user

We are requesting this stays open as a distinct customer escalation. Closing it as a duplicate without addressing the underlying problem only reinforces the lack of communication that #41930 already called out.

👎 on the auto-close.

[tayaee]

Hi OP,

Do you use --continue/-p or --resume/-p a lot? I also have the token drain issue from 2.1.91 on Windows (Claude Pro $20 plan). I wonder if my situation is the same as yours.

Below is for Claude Code team:

The results from 2.1.34 and 2.1.69 are good, but the result from 2.1.91 is bad.

1. Test result with CLI v2.1.34 (Stable/Safe Version): Prompt caching works as expected. The --continue flag successfully hits the cache with minimal misses.

winget install Anthropic.ClaudeCode --version <version_number> --force
npm install -g @anthropic-ai/claude-code@<version_number>

Interactive Session:
/path/to/claude.exe --model sonnet --dangerously-skip-permissions
hi => cache miss 1593 tokens (cache hit 92%)
hi => cache miss 153 tokens (cache hit 99%)
hi => cache miss 72 tokens (cache hit 100%)

CLI Execution with --continue:
/path/to/claude.exe ... --continue -p hi => cache miss 1616 tokens (cache hit 92%)
/path/to/claude.exe ... --continue -p hi => cache miss 156 tokens (cache hit 99%) [Good]
/path/to/claude.exe ... --continue -p hi => cache miss 153 tokens (cache hit 99%) [Good]

2. Teset result with v2.1.69 (Regression Identified): Cache hit ratio drops significantly when using the --continue flag.

winget install Anthropic.ClaudeCode --version <version_number> --force
npm install -g @anthropic-ai/claude-code@<version_number>

Interactive Session:
/path/to/claude.exe --model sonnet --dangerously-skip-permissions
hi => cache miss 3162 tokens (cache hit 83%)
hi => cache miss 58 tokens (cache hit 100%)

CLI Execution with --continue:
/path/to/claude.exe ... --continue -p hi => cache miss 16153 tokens (cache hit 0%)
/path/to/claude.exe ... --continue -p hi => cache miss 1653 tokens (cache hit 90%) [No good]
/path/to/claude.exe ... --continue -p hi => cache miss 1669 tokens (cache hit 90%) [No good]

3. Test result with v2.1.91 (Latest as of April 3rd): The issue persists in the latest version with consistently low cache hit ratios during continued execution.

npm install -g @anthropic-ai/claude-code@<version_number>

CLI Execution with --continue:
/path/to/claude.cmd ... --continue -p hi => cache miss 5250 tokens (cache hit 68%)
/path/to/claude.cmd ... --continue -p hi => cache miss 5260 tokens (cache hit 68%) [No good]
/path/to/claude.cmd ... --continue -p hi => cache miss 5268 tokens (cache hit 68%) [No good]

The full paths to the claude.exe or claude.cmd were:

• %LOCALAPPDATA%\Microsoft\WinGet\Links\claude.exe --version
• %APPDATA%\npm\claude.cmd --version

I used my own tool to monitor the cache miss and hit ratio during the experiments:

• uvx -q ccctail -f

[apolopena]

This is related as it is a service degradation and all of this is not acknowledged by Anthropic at all.
#47886