Welcome to Jumble, your go-to source for AI news updates. This week, Grok 4.6 pulled level with the best models on the market at roughly half the price. Meanwhile, Claude quietly started stamping every word it writes with an invisible mark you cannot turn off. Let’s dive in ⬇️
In today’s newsletter:
⚔️ Musk's model draws level with the leaders
🖋️ Your chatbot now signs its work
🎬 A feature film made for two million
🏁 One frontier model gets fourteen times faster
🧭 Weekly Challenge: stop babysitting your chatbot
🛰️ Grok 4.6 Closes the Gap and Undercuts Everyone
SpaceXAI shipped Grok 4.6 on Wednesday, built for agents that stay on a job across many steps. It scored 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol and trailing Fable 5 by a point.
📊 Frontier Scores at Half the Sticker
API pricing starts at $2 per million input tokens and $6 per million output, well under what rivals charge at that level. The company says the model now checks its own work before moving on.
Credit: xAI
🤖 Bots That Sign Into Your Apps
A day earlier came Grok Bot, agents with their own cloud computer that sign into your tools and work around the clock. It landed alongside the completed Cursor acquisition, and 4.6 powers both.
🕵️ Claude Now Leaves a Mark on Everything It Writes
Anthropic confirmed it is watermarking Claude's text and files to satisfy the EU AI Act transparency code that took effect August 2. It applies at the model level, so it is there whichever product you used.
Will you use AI that watermarks your text or those that give you the option?
📎 It Survives Copy and Paste
Because the mark lives inside the text itself, it travels when you paste and may survive light editing. Files get the C2PA standard instead, and there is no toggle to switch it off.
😤 The Backlash Says the Quiet Part
The loudest complaints came from people worried about getting caught at work or in class, a fairly revealing objection. Google went the other way this week, making its visible watermark optional.
Weekly Scoop 🍦
🔐 Weekly Challenge: Hire a Bot and Give It a Real Job
Challenge: Every lab is selling agents that stay on a job for hours. Most of us still feed our chatbots one sentence at a time, so try the opposite this week.
Here's what to do:
🖥️ Step 1: Pick a job, not a question Choose something with a real finish line you would normally spread across a whole afternoon, like sorting a messy spreadsheet or planning a trip end to end.
👀 Step 2: Write the brief you would hand a person In one message give ChatGPT, Gemini or Claude the goal, the constraints, what done looks like, and permission to make reasonable calls without asking.
🌙 Step 3: Sit on your hands Let it run the whole way through without correcting it, and keep a tally of every moment you wanted to jump in but did not.
🔒 Step 4: Grade the gap Compare the finished thing against what you would have made yourself, then look at your tally and ask how many of those interruptions were actually needed.
If four labs are all within a point of each other, does the leaderboard still mean anything? And would you use Claude less knowing every sentence carries its signature? See you next time! 🚀
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
