Before this dropped, I had basically given up on Claude. 4.7 was so frustrating I'd moved most of my coding to GPT-5.5. Opus 4.8 pulled me right back in. I spent the last few days doing nothing but testing it side-by-side with 4.7 on real builds — here's the honest read. The frame nobody is connecting: - 4.7 dropped April 16. 4.8 dropped 42 days later — fastest Anthropic has ever turned a model around - The reason: they had a compute crisis. Anthropic grew 80x this year (planned for 10x). CEO confirmed it on record - The fix: a $1.25B/month deal with xAI for 220k+ GPUs. The same week the deal landed, Anthropic doubled limits in Claude Code - And the kicker — the price didn't move. 4.8 still $5/$25 per M tokens. Fast mode 3x cheaper What's actually different from 4.7: - The second-guessing loops are gone. 4.7 would talk itself out of the right answer and back into it while you watched tokens drain. 4.8 commits - Long-running tasks finally work. 4.7 drifted, redid old decisions, asked "should I keep going?" every few minutes. 4.8 holds the thread. You can leave it alone on a big job - /goal, dynamic workflows, ultracode — Anthropic shipped the long-run tooling alongside it. The model + the tools finally pull in the same direction Two real tests in the video (built same prompt on both): - Solar system simulator — 4.7 built a working orbit demo. 4.8 added click-a-planet stats, fly-through, and land-on-surface modes. Same prompt - 2001 + 2026 encyclopedia — 4.7 built two good versions. 4.8 tied them together with a single idea: "a clickable link was born blue in 1991 and we kept the blue" Where GPT-5.5 still wins: Terminal-Bench (78 vs 74). Anyone telling you 4.8 sweeps the board is lying. But on Agentic coding it's 69 vs 58 — and it's the strongest model Anthropic has ever shipped on computer use. My verdict: if you gave up on Claude over the last 6 weeks, this is the one to come back for. It's what 4.7 should have been the whole time. Feels like 4.6 with the rough edges smoothed off.

I set up Claude Cowork as a "Data Entry Clerk" — drag in a stack of invoices and receipts, say "log all of these," and it fills in your spreadsheet automatically. Vendor, amount, date, category. Zero typing. 60 receipts logged in ~10 minutes. It even caught a $400 charge I got billed for twice. Watch the video here:https://www.instagram.com/p/DZTX9CIAAno/ Full step-by-step guide (including the exact master prompt) is here: 👉 https://flicker-celestite-7b6.notion.site/The-Data-Entry-Clerk-Log-Every-Receipt-Invoice-Into-Your-Spreadsheet-Claude-Cowork-378d180d8c8081719e22dcb5e3f32e09?source=copy_link Let me know if you set it up — curious what it catches in your books.

Two job postings I found pay up to $500K and $240K a year — and they both want the SAME skill: building AI agents. 90% of developers don't have it yet. Watch the video here:https://www.instagram.com/reel/DZOPDR5AYbG/?utm_source=ig_web_copy_link&igsh=MzRlODBiNWFlZA== The good news? You can learn it for free, even with zero AI background. Go through these in order: 1️⃣ Hugging Face — Agents Course 🔗 https://huggingface.co/learn/agents-course Fully structured, beginner-friendly. Takes you from absolute zero to building multi-agent systems. You finish with a FREE certificate you can put straight on your LinkedIn. → Start here to build your foundation. 2️⃣ Google — ADK (Agent Development Kit) Crash Course 🔗 https://google.github.io/adk-docs/ Teaches you to build agents on Google's own framework. Because Google built it, companies are about to adopt it everywhere — learning it now puts you ahead of the curve. → Do this second to learn a production framework. 3️⃣ LangChain Academy — Deep Agents 🔗 https://academy.langchain.com/ Covers LangGraph and the Claude Agent SDK — the exact tools those high-paying job postings are asking for. Tons of companies already run this in production. → Finish here to learn what employers want right now. 📌 The Order That Works: Foundation (Hugging Face) → Framework (Google ADK) → Production Tools (LangChain). Do all three and you'll go from zero to genuinely employable in one of the most in-demand skills on the market. Of course, join our AI Accelerator+ to learn more beyond that :) Drop a 🔥 if you're going to start this week.

This isn't a leak. It's not a rumor. Anthropic just published internal data on how much of Claude's development is now being done by… Claude. The numbers are nuts. Read these slow 👇 → 80%+ of all code merged into Anthropic's codebase is now written by Claude. (Before Claude Code launched in Feb 2025, that number was in the low single digits.) → Anthropic engineers now ship 8x as much code per quarter as they did from 2021–2025. → On a standard test — "take this small AI training script and make it faster" — a skilled human takes 4–8 hours to reach 4x speedup. Claude Opus 4 (May 2024) averaged ~3x. Mythos Preview (April 2026)? ~52x. → When shown a research session where a human made a wrong turn, Mythos Preview picked a better next step 64% of the time — up from 22% in 2024. → One April 2026 example: Claude shipped 800+ fixes that cut a class of API errors by 1,000x. Engineering estimate of how long that would've taken a human: four years. Anthropic's own framing: this is a possible path to recursive self-improvement — AI autonomously building a more capable successor. They say it's not here yet and not inevitable. But it's "happening faster than we thought." My read on what this actually means: The "AI builds AI" headline is the dramatic part. The part you should actually internalize is the rate of acceleration itself. Look at the task-length curve they shared. Opus 3 (Mar 2024) could reliably handle 4-minute tasks. Sonnet 3.7 (a year later) — 90 minutes. Opus 4.6 (a year after that) — 12 hours. Mythos Preview can hold a session for at least 16 hours and that's the top of what they can currently measure. Task length is doubling roughly every 4 months. That used to be every 7. The curve is bending. What this means for builders: Whatever you couldn't automate 6 months ago — you can now. The "the AI can't hold a long enough chain of thought to do this end-to-end" objection is dying in real time. Workflows that needed three orchestrated agents last quarter can be one agent in 90 days. Plan accordingly.

Most people use Claude Code completely wrong. The defaults teach you to type into one chat, hit enter, wait, type again. That's the loop. The guy who actually built Claude Code does basically the opposite. Boris Cherny runs ~15 Claude Code sessions at the same time, barely types most of his prompts, and dispatches work like a team lead. I read every public post he's made and put the whole thing on one page. The 7 moves: 1. Parallel sessions — 5 in terminal + 5–10 in browser. One session, one job. Opus + thinking mode for all of them 2. CLAUDE.md — the rules file Claude loads automatically. Every time it gets something wrong, you write the fix in. Compounding engineering — month 3 has 50 rules nobody else has 3. Plan mode — Shift+Tab+Tab. Pour energy into the plan, then auto-accept the implementation 4. Slash commands + sub-agents — .claude/commands/ + .claude/agents/. Boris uses /commit-push-pr, /simplify, /verify, /go dozens of times a day 5. /loop — local background loops up to 3 days. His 4: /babysit, /slack-feedback, /post-merge-sweeper, /pr-pruner 6. /schedule — same idea, runs in the cloud. Laptop can close. GitHub events can trigger it 7. Verification (the #1 tip) — "grill me on these changes", "prove it works", "scrap this and implement the elegant solution" The adoptable system if you're not a developer: 2–3 parallel sessions (not 15) + CLAUDE.md + plan-first discipline + named prompts for anything you do twice + 1 loop during workday + 1 schedule overnight + a verification line on every important task. If you do only ONE thing from the video: add a verification line. The quality jump is immediate. 📎 Full Boris Cherny guide PDF pinned below — every move with the setup steps + the Keep-vs-Skip cuts for non-devs. Comment which move you're copying first 👇 [Watch the video here ▶️]