Three AI image models. 15 rounds. Same prompts. One of them you can download and run on your own GPU. GPT Image 2, Nano Banana 2, and open weights Ideogram 4, head to head on photorealism, text, spatial reasoning, and style. The closed giants were supposed to win easy. They didn't. The open one came in third by a hair, then beat both giants at rendered text, the thing AI is usually worst at. You don't need anyone's API for top tier results anymore. Full scorecard at the end. Here's the link to get started with Ideogram 4.0 👍 https://docs.google.com/document/d/1vveMIlHaM12fNERyfSzAtvQiyFSfG4cYvGqgXgxtWlk/edit?usp=sharing

Meet GPT-5.6 Sol, Terra, and Luna, OpenAI's new flagship lineup, and they launched under direct coordination with the US government before anyone outside ~20 partner orgs could touch them. - Sol is the flagship. Built for deep research, complex coding, and agentic work that runs for hours unsupervised. - Sol also has "Ultra Mode," which spins up multiple sub-agents working in parallel on the same task. - On cybersecurity benchmarks, Sol matches Anthropic's Mythos Preview using roughly a third of the output tokens. That's a real efficiency jump. - Terra is the mid tier. Performance close to GPT-5.5, at about half the price. - Luna is the speed and volume layer. Fast, cheap, built for high frequency lightweight tasks. - Here's the catch. None of this is publicly available yet. No waitlist, invite only, roughly 20 partner organizations. - This is reportedly the first US frontier model to launch under a government managed access list, which is a new thing in AI releases. - Broad access is "coming weeks" with no confirmed date. If you're building with AI right now, this is worth watching closely. Not just because of what these models can do, but because it might be a preview of how frontier AI gets released going forward, government checkpoint first, public access second. So I'm curious, if you had to pick your tier today, are you a Sol person running long agent workflows, a Terra person who just wants solid performance without burning budget, or a Luna person doing high volume lightweight stuff? As for me, I'll be with Terra. Link: https://openai.com/index/previewing-gpt-5-6-sol/

Open source just caught the frontier. GLM-5.2 sits 1% below Claude Opus 4.8 on real-world coding, beats GPT-5.5, and costs 1/6 the API price. Here's the link to get started with GLM 5.2 👍 https://docs.google.com/document/d/1q8WAqnyl-tE5r0qX8gvm3wfI6xXg0KbkOZCeEbwWfV4/edit?usp=sharing

Your Slack just got a whole lot smarter. Anthropic just launched @Claude Tag, and it's basically like adding an AI teammate directly into your Slack channels. - Tag @Claude in any thread and it reads the full context, then does the work - Can catch up on long messy threads and surface decisions, open questions, and action items in seconds - Pulls live data like top 20 accounts by spend for the last 7 or 28 days right into the channel - Turns a bug report into a draft PR without leaving Slack - Set standing instructions and it watches channels, triages alerts, and only tags you when a human decision is needed - Connects to GitHub, Linear, Asana, Salesforce, HubSpot, Snowflake, Figma, Zendesk, and more - Has its own account identity in your tools so every action is logged and traceable - Available now for Claude Enterprise and Team plans in Slack This is AI moving from "chat assistant" to "actual team member." The fact that it can run long background tasks, monitor channels proactively, and tag you back when something needs a decision is a totally different level of agentic automation. Drop a comment if your team is already in this or thinking about it. Read More: https://support.claude.com/en/articles/15594475-what-is-claude-tag Set Up Instructions: https://claude.com/docs/claude-tag/admins/setup-overview

Looks like you're not going to be working out of Antigravity as your primary agent anymore. Like me, it autoupdated to 2.0 today and 2.0 is not an upgrade, it's a new product and it's not an IDE. Back to VS Code I suppose. Interested what everyone else is using.