Caveman, the install that shortens what Claude writes back, took one of my tasks from 1180 tokens to 159, which is about 87 percent, and that number is real. I want to be careful with it, though, because that was my best result. Other tasks came back only 41 percent and 30 percent smaller. So the honest range from my own runs is 30 to 87 percent depending on the task, and the only way to know yours is to test it on something you repeat. Keep in mind that output is a smaller share of your bill than input. Even a large percentage off the smaller slice saves less in absolute terms than trimming what goes in, which is why rtk, the install that works on the input side, is the bigger win of the two. Try it today: pick one task you run often, note the token count of the reply, install caveman, run the same task again and compare the two. Which task do you run in Claude Code most days, and do its replies tend to be long enough that shortening them would matter?

Select your answer in the poll 👇

n8n 2.44.0 refuses to save a workflow that adds or changes a Function or Function Item node, and the Code node is not affected. The two look alike on the canvas and both hold JavaScript, so check which one you actually have before you rebuild anything. The Function node ran your code once and handed you every item in an array called items. Function Item ran once per item and handed you that item's data as item. The Code node replaced both back in n8n 0.198.0, and n8n hides the old pair from the node picker, so anything you added from the picker lately is already a Code node. The old ones live on in older workflows and in exported JSON, still running in vm2, a sandbox n8n calls insecure. Nothing stops running today. The save is what fails: add one, edit one, or import a JSON file that has one. Open the node, and if its panel says Function, replace it with a Code node set to Run Once for All Items, where items becomes $input.all(). If it says Function Item, use Run Once for Each Item, where item becomes $json. Paste the old code in, change those names, and run it once against the same input. It is a beta release, so there is time. Self hosters can set N8N_DEPRECATED_NODES_BLOCK=false while they move. 🔗 https://github.com/n8n-io/n8n/pull/31360 Does any node in your oldest workflow still say Function in its panel?

An Anthropic hackathon winner open sourced his entire Claude Code setup, ECC, and it has passed 270k stars. Installing all of it in one go is what maxes out your credits, and the subagents are not the reason. The drain is every connected tool switched on at once. 🔗 The repo: ➡️ https://github.com/affaan-m/ECC I tested it, and three pieces are worth having. The guide attached to this post installs only those three by hand and leaves everything else in the repo, so nothing else from it touches your context. 💡 The three tools ✅ Memory: learns from your sessions and reuses your last bug fixes, so Claude Code stops falling for the same bug ✅ Planner: writes a detailed plan and waits for your yes before a single line of code ✅ Budget: shows how full your context is and which installed tools are draining it 📎 The free file is attached to this post ➡️ ecc-three-tools.zip The exact copy commands, the one hook the memory skill needs, and how to use each tool once it is in. Which part of your setup do you suspect is eating your tokens? Tell me in the comments. Book a free Consultation: 🔗 https://go.genovaflow.com/skool-ecc-without-the-bloat-3-claude-code-tools-worth-call Sponsorship: 📧 [email protected]

Google opened Gemini 4 Argon to everyone on 13 October, and on Google's own benchmark table it leads Claude Opus 5.5 on automation work while Opus still wins in the terminal. These are Google's numbers, Argon first and Opus 5.5 second: - AutomationBench: 51.3% against 42.5% - DeepSWE v1.1: 77.9% against 74.2% - Terminal Bench 4.0: 57.4% against 66.4% - FrontierSWE v2: 55.0% against 62.3% It is in the Gemini API through AI Studio now, at an introductory $2 per million input tokens and $10 per million output, going to $4 and $20 after 30 November. The Gemini app has it for Pro and Ultra subscribers. The first hands on reports are more mixed than the table. ariwilson on Hacker News ran his own eleven bug fixes through it and it solved ten, but a median run cost him $0.52 against $0.18 for Sonnet 5.5 and took about four minutes against 32 seconds, because it wrote around 24k output tokens where Sonnet wrote about 3k. eyalitki gave it a payroll question, watched it get every edge case right, and then saw it work out 430 minus 60 as 410. Google's own engineers left it running overnight on ADK, their 200,000 line Python agent framework, and the changes it made are now live, with in memory session prep in long sessions down 40 percent. If you build with n8n or Claude Code, I would not move anything yet. An automation step that keeps failing on a long, many step task is the cheap thing to try it on at the introductory price, and for work in the terminal Opus 5.5 still has the better number. 🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ The tester's numbers: https://news.ycombinator.com/item?id=50077034 Which step in your own workflows would you hand to a new model first?