Why Claude Keeps Hitting Usage Limits (& how to fix it)

Concise summary

Claude now enforces 5-hour “current session” usage limits more strictly during peak hours. As a result, chats can stop mid-task when the session allowance is exhausted, even if your weekly quota remains.

Problem and recent change

• Claude enforces a 5-hour “current session” usage window more strictly during peak hours to manage demand.
• Peak windows: ~5:00–11:00 AM PT (1:00–7:00 PM GMT).
• Weekly limits are unchanged, but you’ll consume the 5-hour allotment faster during peak times — causing chats to stop mid-task when the session allowance ends.

Key technical concepts

• Session usage vs length (context) limits
  • Session: a 5-hour usage window that starts when you send your first message; tracked in Settings → Usage.
  • Length/context window: the amount of text Claude can hold/process in a single chat (measured in tokens). Most plans: ~200k tokens per chat; some plans up to 1M.
• Tokens
  • Internal unit (roughly 0.5–0.75 words per token).
  • Every message, response, file, and tool call consumes tokens.
• Model differences
  • Opus 4.6 (higher-end) burns tokens about 5× faster than Sonnet/Sonic 4.6 (more efficient).
  • Use Opus only for heavy tasks when you have spare session budget.
• Tools/connectors/plugins/MCPs/skills
  • These can consume context tokens even when idle.
  • Some connectors (e.g., Apify, Chrome integrations, Drive/Notion/Gmail) are major token sinks.
• Useful commands/features
  • /context command: reveals per-chat token breakdown (system prompt, tools, skills, files, memory) so you can see what’s burning tokens.
  • Extended Thinking: a feature that increases internal computation and token use. Toggle it off when not needed.
  • Extra usage/credit: you can add small credit (pay-as-you-go) to finish tasks when session limits are hit.

Practical fixes — 14 tips

1. Monitor usage: check Settings → Usage to see session and weekly usage and reset timers.
2. Use /context: inspect what’s consuming tokens in the chat (system prompt, tools, MCPs, files).
3. Choose models wisely: default to Sonnet/Sonic for efficiency; use Opus for intensive jobs when you have spare session budget.
4. Turn off Extended Thinking when unnecessary.
5. Disable unused tools/connectors/MCPs/plugins (especially duplicate connectors like multiple Chrome integrations).
6. Start a new chat per task: avoid long mixed-topic threads that keep accumulating context.
7. Be specific in initial prompts: tell Claude exactly what to do to avoid extra back-and-forth and unnecessary processing.
8. Batch requests into single prompts: group multiple edits/asks together to reduce repeated overhead.
9. Preprocess heavy documents: convert PDFs to markdown/plain text before ingestion (markdown is LLM-friendly and lighter).
10. Use other services for extraction: e.g., Perplexity or other tools to convert/extract content before loading into Claude.
11. Use Claude Projects / RAG: projects can retrieve only relevant snippets instead of loading all files into the chat.
12. Trim custom instructions & keep them short (recommend <500 words): every chat load reads instructions and costs tokens.
13. Remove pointless/idle files from projects/co-work folders to avoid accidental extra context processing.
14. Build and use Claude Skills for repeatable workflows: skill descriptions load minimal context and encode repeatable recipes so Claude doesn’t re-learn steps.

Operational tricks

• Session reset trick: send a throwaway prompt earlier to shift the start of the 5-hour window so you can better align heavy work with non-peak periods (effectively get two usable sittings).
• Schedule heavy jobs outside peak hours so tokens/session go further.
• Keep Projects clean, use short instructions, and leverage memory where helpful to avoid re-explaining.

Takeaways

• The apparent “sudden” drop in capacity is due to stricter peak-hour session throttling combined with token consumption from models, tools, and context.
• Quick wins: switch to Sonic (Sonnet), toggle off Extended Thinking, remove unused connectors, use /context to find token hogs, batch prompts, and preprocess files.
• Longer-term: move heavy ingestion into preprocessed markdown or Projects/RAG, and build Skills for repeatable tasks.

Relevant product features & tools mentioned

• Claude/Anthropic: session usage tracking, context windows, models (Opus vs Sonnet/Sonic), Extended Thinking toggle, Skills, Projects, Co-work, memory, /context command.
• Connectors/plugins: Drive, Gmail, Notion, Apify, Chrome integrations, MCPs.
• External tools: Perplexity (for PDF → markdown extraction).

Main speaker / sources

• Primary speaker: an unnamed video presenter — a Claude/AI power-user explaining the changes and giving a 14-point guide.
• Sources referenced: Claude update posts (Reddit/posts about session limits), Anthropic/Claude product settings and features, and external tools like Perplexity and Apify.