Anthropic raised the thresholds this week. Here is what actually changed, what did not, and how to make your Claude session go further. **What changed (May 06):** - Claude Code 5-hour session limits **doubled** for Pro, Max, Team, and Enterprise. - Peak-hours throttle on Claude Code **removed** for Pro and Max. - Claude Opus API rate limits raised "considerably." - Powered by a new SpaceX compute deal: 300+ MW of capacity, around 220,000 NVIDIA GPUs. **What did NOT change:** - Weekly limits. If you were hitting the weekly cap last month, you will still hit it this month. - The change buys you more in a single session, not more across a week. --- ## 5 ways to stretch your Claude session (desktop users) These work on Claude.ai and the desktop app. No code required. **1. Use Projects, stop re-pasting the same context.** If you keep pasting the same brief, doc, or code file into every new chat, move it into a Project once. Every conversation inside that Project sees those files automatically. You stop burning your session limit re-uploading the same information, and Claude stops re-reading it. **2. Start a new chat when the topic changes.** Every message in a long chat sends the entire history back to Claude. After 30 or 40 turns, you are paying (in session usage and in speed) for the first message every single time. When you switch topic, start fresh. **3. Pick the right model for the task.** Haiku for quick lookups, summaries, and short questions. Sonnet for normal day-to-day work. Opus only when the task is genuinely hard (architecture, dense reasoning, long multi-step analysis). Most people leave Opus on for everything and burn through the limit much faster than they need to. **4. Be specific in your first message.** Vague prompts trigger 5 rounds of clarification. One precise prompt with full context and constraints lands the answer in one shot. Spend 30 seconds writing the prompt, save 10 minutes of back and forth. **5. Use Artifacts instead of asking for full rewrites.**

Tokens = Money. Most people burn 3–5× more than they need on Claude without realizing it. I put together a 2-page cheat sheet covering everything I wish someone had handed me when I started: Page 1 — the essentials - Pricing per 1M tokens (Opus, Sonnet, Haiku) - The Savings Stack (these multiply, not add — 80–95% cuts are normal) - The 5 core tactics: right-sizing the model, prompt caching, shorter prompts, output control, context trimming - Common token wasters to stop doing today Page 2 — the deep dive - Claude Code in the terminal: slash commands, CLI flags, how to slim your CLAUDE.md - Prompt caching full Python example (what to cache, what the response tells you) - Prefill tricks that kill the "Sure, here you go..." preamble - Model decision tree (when Haiku, Sonnet, Opus each) - Real-world recipes: cheap classifier, long-doc Q&A, cost-capped agent loops, batch evals - A pre-ship checklist you can run through before deploying 📎 Attached: Token_Saving_Cheat_Sheet Drop a comment with the one tip that surprised you most, or your own token-saving trick. Vincent