Hello everyone,
Today I will be at OpenAI DevDay. If you see me there, feel free to stop by and say hello—I’ll be wearing pink shoes.
Sam mentioned they’ve “discovered something new“.
While I’m under embargo regarding all the upcoming launches, I’m bound by secrecy and would love to be invited back, so I can’t share details just yet. However, there are some rumors circulating on X, and it seems a few things align with our expectations.
Later today, I’ll also be giving a talk at Stanford. Unfortunately, I’ve been awake since 2 a.m. 😩—I’ve never had issues with jet lag before, so what a day for it to hit! 🙃
Only in San Francisco… Last night at the hotel bar, someone beside me asked about the tool I was using. I proudly replied, “I built it—it’s the course platform I created.” Now I have a new friend at DeepMind!
Ben’s Bites is supported by Google Cloud
Google Cloud kicks off Season 3 of Advent of Agents this Thursday, October 1. You’ll receive one question a day about effectively running agents in production—covering topics from agent identity to kill switches, complete with a short video and code for you to build with your coding agent. Get Day 1 straight to your inbox.
Anthropic has launched Claude Sonnet 5.5 – Sonnet 5.5 shows promising performance based on benchmark results, bringing it close to Opus 5.5 in coding while significantly enhancing its capabilities in understanding images, charts, and more. If you’re subscribed to the $20 plan, you might want to try replacing some of your typical requests to Opus with Sonnet 5.5 for complex tasks.
Sonnet used to be my go-to. I relied on it until I switched to Opus 4.5, but after seeing numerous positive reviews for this new version, I’m eager to give it another try—just not on Claude Code; I find their harness challenging to navigate.
By the way, Anthropic consistently delivers high-quality .5 versions of their Sonnet models.
Introducing Manus 2.0 and Cue. Manus is back with a video editor, a game builder, and the usual automations (through email, Slack, calendar, etc.). Cue focuses on consumer agents: each agent gets a dedicated email, phone number, wallet, and computing resources, similar to Grok Bot, Muse, and others. It’s currently free during the early access phase (use code MEETCUE if you’d like to try it out).
GPT-6 Astra excels in utilizing a computer or browser. I recently asked Opus 5.5—my daily tool—to perform some tasks in Chrome, and I found it lacking compared to Astra (or any GPT model, to be fair). Here are some examples of how it can be beneficial. OpenAI also fixed a bug that had been hindering GPT-6 Sol and Luna’s image comprehension capabilities.
OpenAI’s agents have been “breaking out” and “hacking” websites. Recently, they managed to peek into a few Australian government sites. Media coverage can sometimes exaggerate the situation, but if you’re interested, here’s the official link to keep up with the actions of these agents.
The latest Meta Enterprise Platform aims to sell Muse and Meta’s AI APIs to businesses. In an interesting turn, Zuckerberg tapped MongoDB’s CEO to lead this new venture, which resulted in a 20% drop in MongoDB’s stock.
* sponsors who make this newsletter possible 🙂
Interested in partnering with us for next quarter?
Email us at shanice@bensbites.com or k@bensbites.com


