Claude Usage Limits: Why They're Tighter & 14 Ways to Fix It

Claude reduced session capacity during peak hours

Anthropic announced adjustments to the 5-hour session limit during peak demand times (5–11 AM PT / 1–7 PM GMT) to manage growing user load. Weekly limits remain unchanged, but you burn through your 5-hour session faster during these windows—similar to surge pricing on public transport.

Peak-hour throttling window (Anthropic announcement)

Session usage limits reset every 5 hours

Your 5-hour session window begins the moment you send your first message to Claude. Within that window, you have a capped allowance of work (tokens) you can do. Once exhausted, you must wait until the window resets or pay for extra usage. Weekly limits are separate and remain unchanged.

5 hours

Session window duration from first message

Session resets every 5 hours; peak hours consume allowance faster

Context windows hold 200k tokens per chat (usually)

Each chat has a context window—the amount of information Claude can hold and process at once. Most plans include 200,000 tokens per context window; some can reach 1 million. Tokens are roughly 0.75–0.5 words each. Every message, file, and response consumes tokens from this pool.

200,000

Tokens per context window (standard)

Can scale to 1M tokens on some plans; every message/file/response consumes tokens

Model choice dramatically affects token burn rate

Different Claude models consume tokens at different rates. Opus 4.6 burns tokens 5× faster than Sonnet 4.6. Haiku is cheapest but may require more back-and-forth trips, making it less efficient for complex tasks. Sonnet offers the best balance for most everyday work.

Token burn rate by model (Opus costs 5× more than Sonnet)

Use the /context command to audit token usage

Type '/context' in Claude to reveal a hidden dashboard showing exactly what is consuming your tokens: system prompts, MCPs, connectors, tools, memory files, and skills. This reveals hidden drains like duplicate Chrome extensions or unused tools running in the background.

Disable extended thinking unless needed

Extended thinking mode forces Claude through extra reasoning steps and outputs, burning significantly more tokens. Turn it off by default and only enable it for genuinely complex tasks. This is one of the biggest token drains people overlook.

Disabling extended thinking saves substantial tokens per session

Remove duplicate or unused connectors and MCPs

Tools like web search, drive connectors, and MCPs run in the background consuming tokens even if unused. Audit your connectors (via the + button) and disable duplicates or tools you don't actively use. Example: having both old Control Chrome and new Claude-in-Chrome MCP wastes tokens.

Switch to Sonnet for everyday tasks

Default to Sonnet 4.6 instead of Opus for routine work. Reserve Opus only for complex, intensive tasks where the extra capability justifies 5× token cost. This alone can dramatically extend your session allowance.

Start a new chat per task or topic

Long chat histories accumulate as overhead—every prompt you send must re-process all previous messages. Switching to a fresh chat when changing topics (e.g., from spreadsheet analysis to marketing copy) eliminates this context drag and resets your token baseline.

Be specific in prompts to avoid wasted processing

Vague prompts like 'summarize this document' force Claude to process entire files. Instead, say 'summarize the financial risks in section three of this PDF.' Specificity reduces unnecessary token consumption by narrowing Claude's search scope upfront.

Batch multiple requests into one prompt

Instead of sending three separate messages ('Fix typo in paragraph 2,' 'Shorten intro,' 'Add CTA'), combine them: 'Fix typo in paragraph 2, shorten intro, add CTA.' This eliminates three sets of token overhead and completes the work in one shot.

Batching requests reduces overhead and token burn

Convert PDFs to markdown before uploading

Heavy PDFs can consume up to 80% of a session's tokens just extracting information. Use Perplexity or another tool to convert PDFs to lightweight markdown or plain text first. This preserves content while drastically reducing token weight, then paste the markdown into Claude.

Converting PDFs to markdown can save 50–80% of token overhead

Use Claude Projects with RAG for efficient file handling

Projects use retrieval-augmented generation (RAG), pulling only relevant content into context rather than loading all files. Store markdown files (not heavy PDFs) in projects, add short custom instructions, and enable memory to pick up where you left off without re-explaining.

Keep custom instructions under 500 words

Long custom instructions load with every message, adding token overhead. Trim instructions to essentials—keep them under 500 words, focused, and to the point. Every word loaded is a token burned on every prompt.

500 words

Maximum recommended custom instruction length

Shorter instructions = lower token overhead per message

Remove idle files from projects and workspaces

Unused files in projects or co-work folders add clutter and potential confusion for Claude, forcing it to filter through irrelevant content. Delete files you no longer need to keep context clean and focused.

Build Claude Skills for repeatable workflows

Skills encode step-by-step processes (e.g., invoice generation, report formatting). Only the skill description loads into context by default, not the entire skill code. This lets you execute complex, repeatable tasks without re-explaining or letting Claude figure out the process each time.

Schedule heavy work outside peak hours

Peak hours (1–7 PM GMT / 5–11 AM PT) throttle session capacity. Plan heavy Claude work for off-peak times (early morning, late evening, weekends) to maximize token efficiency. If using scheduled tasks, set them to run outside peak windows.

Schedule heavy work during off-peak hours for maximum efficiency

Use the session reset trick to extend capacity

Your 5-hour window starts when you send your first message, not when you log in. Send a throwaway prompt early (e.g., 'Hi, let's get started') to start the window, then return 2–3 hours later to begin real work. By the time your first window expires, you can start a second session, effectively doubling your working time.

Session reset trick: stagger prompts to access two 5-hour windows in one day

Pay for extra usage to avoid interruption

If you hit your session limit mid-task, you can add credit to your account (e.g., £8/month) to continue on extra usage rather than waiting 3+ hours for the window to reset. This is useful when you're in the middle of video production or urgent work.

All 14 efficiency strategies at a glance

The video presents a tiered approach: Tier 1 (settings) includes switching to Sonnet, disabling extended thinking, auditing with /context, and removing unused tools. Tier 2 (habits) covers starting new chats per task, being specific in prompts, and batching requests. Tier 3 (workspace) involves converting PDFs to markdown, using projects with short instructions, and building skills. Tier 4 (advanced) includes scheduling off-peak and using the session reset trick.