I stopped hitting Claude's usage limits - 10 things I changed
kaize · tweet · 2026-04-10 Reader 🖥️
You can avoid hitting Claude's usage limits by understanding that Claude counts tokens (not messages) and adopting ten habits that minimize unnecessary token consumption: editing instead of following up, starting fresh chats, batching questions, caching files in Projects, saving preferences to Memory, disabling unused features, choosing cheaper models for simple tasks, spreading work across the day, avoiding peak hours, and enabling overage billing.
practical · listicle / how-to guide — A practical advice thread on optimizing token usage when working with Claude. Written for everyday Claude users who hit rate limits. No theoretical framework beyond a simple cost formula.
Structure
| Part | Role |
|---|---|
| Introduction (framing the problem) | Establishes the core insight: Claude counts tokens, not messages, so the solution is to use tokens wisely. |
| Tips 1-3 (conversation-level habits) | Address how you structure your interactions within a single conversation to reduce token waste from context re-reading. |
| Tips 4-6 (platform features) | Show how Claude's built-in features (Projects, Memory, feature toggles) can eliminate repeated token spend across conversations. |
| Tips 7 (model selection) | Introduces model-tier awareness as a token-budget strategy. |
| Tips 8-9 (temporal strategies) | Explain how Anthropic's rolling-window and peak-hour pricing mean timing matters. |
| Tip 10 (safety net) | Provides a fallback for when all other strategies are insufficient. |
| Conclusion | Reinforces the core message: once habituated, these practices make limits a non-issue. |
Key terms
- tokens
The unit Anthropic actually meters for rate-limiting and billing — distinct from 'messages' in the author's framing. - context load / context re-reading Context Rot & Context Limits
The process by which Claude re-processes the entire conversation history with each new turn, multiplying token cost. - rolling 5-hour window Usage Limits & Rationing
Anthropic's rate-limit accounting period: usage older than 5 hours no longer counts against your current limit. - peak hours Usage Limits & Rationing
Weekday mornings Pacific Time when Anthropic applies a higher token-cost multiplier against your session limit.
Propositions
- explicitClaude counts tokens, not messages, for rate-limiting purposes.conf: high
- explicitToken cost per conversation turn grows quadratically with message count because the full history is re-read each turn.conf: high
- explicitEditing your original prompt and regenerating replaces the old exchange rather than stacking a new one, saving tokens.conf: high
- explicitStarting a new chat every 15-20 messages prevents runaway token costs.conf: high
- explicitIn a 100+ message chat, approximately 98.5% of tokens are spent re-reading history and only 1.5% on generating new output.conf: medium
- explicitBatching multiple questions into one prompt saves tokens and often produces better answers because Claude sees the full picture.conf: high
- explicitFiles uploaded to Projects are cached and do not re-consume tokens on repeated access.conf: medium
- explicitWeb search, connectors, Explore mode, and Extended Thinking all add token overhead even when not needed.conf: medium
- explicitHaiku can handle simple tasks at 50-70% lower token-budget cost than Sonnet or Opus.conf: medium
- explicitStarting March 26, 2026, Anthropic applies a higher rate-limit cost multiplier during weekday peak hours (5-11 AM PT).conf: high
- explicitPro and Max subscribers can enable overage billing at API rates to avoid hard blocks.conf: high
Arguments
You should edit your prompt instead of sending follow-up corrections. strength: high
Explicit premises:
- Claude re-reads the entire conversation history every turn.
- Each follow-up adds to that history, increasing cost quadratically.
- Editing replaces the exchange rather than stacking.
Implicit premises:
- The user's goal is to minimize token usage.
- The edit-and-regenerate feature is available on Claude's interface.
Inferences:
- Editing avoids the quadratic cost growth that follow-ups cause.
You should start a new chat every 15-20 messages to avoid token waste. strength: medium
Explicit premises:
- Token cost grows quadratically with conversation length.
- In long chats, nearly all tokens go to re-reading history.
Implicit premises:
- The summary-and-restart workflow preserves enough context to continue the task.
- The cost of summarizing + starting fresh is less than the cost of continuing.
Inferences:
- Periodically resetting the conversation keeps you in the low-cost region of the quadratic curve.
You should choose the cheapest model adequate for each task. strength: medium
Explicit premises:
- Haiku handles simple tasks adequately.
- Using Haiku for simple tasks frees 50-70% of your budget.
Implicit premises:
- Users can reliably judge which tasks are 'simple' vs 'complex'.
- The different models share the same rate-limit pool or budget.
Inferences:
- Model selection is a budget allocation decision, not just a quality decision.
Narratives
- The author's personal journey from blaming Claude for strict limits to realizing the limits reflect token accounting that users can optimize.
→ Establishes credibility and frames the reader's likely frustration as a solvable problem. - An unnamed developer who tracked usage in a 100+ message chat and found 98.5% of tokens went to re-reading history.
→ Provides social proof and a concrete data point to dramatize the waste from long conversations.
Questions addressed
- ✅ Why do I keep hitting Claude's usage limits?
Because Claude counts tokens, not messages, and common usage patterns (long chats, follow-up corrections, repeated uploads) waste tokens through context re-reading. - ✅ What concrete habits can I adopt to reduce token usage?
Ten specific practices organized from conversation-level habits through platform features to temporal strategies and billing safety nets. - ⭕ How does Claude's rate-limiting actually work?
Partially addressed — the rolling 5-hour window and peak-hour multiplier are mentioned, but the exact mechanics (how tokens map to limits, how different models count differently) are not fully explained.
Gaps
- unsupported claimThe 50-70% budget savings from using Haiku is stated without evidence or calculation.
Read next: Look for Anthropic's published pricing ratios between Haiku, Sonnet, and Opus to verify. - oversimplificationThe quadratic formula S*N(N+1)/2 assumes equal-length exchanges and ignores that Claude's responses vary greatly in length. It also conflates input tokens (which are re-read) with output tokens (which are not re-read).
Read next: Anthropic's documentation on how input vs output tokens are counted for rate limits. - missing nuanceThe claim that Projects caching means files 'don't eat into your usage' is stated without caveats. Anthropic's actual caching behavior and whether cached tokens truly have zero rate-limit impact is not substantiated.
Read next: Anthropic's documentation on prompt caching and how cached tokens count toward limits. - temporal sensitivitySeveral tips reference specific Anthropic platform features and policies (peak hours starting March 26, 2026; overage billing; model tiers) that may change. The advice is tightly coupled to the current product state.
Read next: Anthropic's changelog and pricing pages for current policies. - missing considerationNo discussion of Claude Code, API usage, or non-web-UI contexts. All advice is implicitly scoped to the Claude web/desktop interface.
Read next: Separate guides for API and Claude Code token optimization.