Key Takeaways
- Max plan users see 3–4x faster limit exhaustion.
- Audit token-heavy tasks, use /compact and CLAUDE.md rules to reduce waste per session.
What Changed — The Regression in Detail
A Reddit user on the Max 5x plan reports a dramatic shift: where they once ran 3–4 Claude Code sessions in parallel for 3–4 hours, they now exhaust the entire 5-hour window with a single session in under an hour. The concrete numbers are stark:
- A single
continuemessage — literally just sending "continue" — consumed ~3% of the 5-hour limit. - One repository audit processing 1.5 million tokens used ~30% of the window.
- Lightweight tasks (updating docs, editing localization strings, small UI adjustments) that previously took ~4 hours now hit the limit almost immediately.
The user also reports unexpected server/content-policy restriction messages on an ordinary manicure/nail application — no sensitive content, no copyrighted material.
This isn't a new project or heavier workload. The user explicitly says: "The workload is considerably lighter than what I was doing several months ago."
What It Means For You — The Likely Culprit
Three possibilities explain this:
Usage accounting changed. Anthropic may have tightened how "active time" is calculated — perhaps counting model inference time differently, or including context loading in the meter.
Internal model costs increased. Claude Code likely uses Claude Opus 4.6 (or newer models) which may be more expensive per token on the backend. If Anthropic didn't adjust the Max plan's effective token budget proportionally, you get less work per session.
A bug in limit tracking. The 3% per "continue" seems disproportionate. If this is a bug, it may affect all Max plan users.
Regardless of cause, the effect is real: you're getting less done per session than before.
Try It Now — Actionable Steps to Stretch Your Limits
Until Anthropic clarifies or fixes this, here's how to maximize your remaining budget:
1. Audit Your Token Waste
Run this in your project to see what Claude Code is loading:
claude code --dry-run --verbose
This shows the full context being sent. Look for:
- Large files being loaded unnecessarily
- Entire directories when you only need specific files
- Repetitive conversation history
2. Use the /compact Flag
Claude Code's /compact flag reduces token usage by summarizing conversation history. Add it to your startup command:
claude code /compact
This can cut token consumption by 40% or more on long-running sessions, according to community reports.
3. Tighten CLAUDE.md Rules
Add a rule to limit context loading. In your CLAUDE.md:
## Token Budget Rules
- When asked to audit or analyze, ask user which specific files first.
- Never load more than 3 files at once unless explicitly requested.
- Summarize what you found rather than reproducing full file contents.
4. Start Fresh for Each Task
Instead of continuing a long conversation, start a new session for each focused task. The continue command's 3% hit suggests conversation history is expensive. Use:
claude code "Implement the login form"
Rather than resuming an existing session.
5. Monitor Your Usage in Real-Time
Install cc-context-telemetry (mentioned in our recent digest) to see context window and rate-limit percentages directly in your terminal:
npm install -g cc-context-telemetry
claude code --with-telemetry
The Bottom Line
This regression is real and documented. If you're on a Max plan and hitting limits faster than expected, you're not alone. The workarounds above can help stretch your budget, but long-term, Anthropic needs to address whether usage accounting, model costs, or a bug is the root cause.
Have you experienced similar regression? Share your data points in the comments below.
Source: reddit.com
