Howdy wizards,
Here’s what’s brewing in AI.
The big thing
OpenAI launched GPT-5.6 and ChatGPT Work, OpenAI’s answer to Claude Cowork.
The model first. GPT-5.6 comes in three versions: Sol (the flagship), Terra, and Luna (the cheap, fast one). Sol performs a hair below Anthropic’s new Fable 5 on overall intelligence, according to Artificial Analysis’ index. And it’s ahead on agentic coding. It’s also token-efficient, and runs at only a third of the cost of Fable 5.
The other launch is ChatGPT Work: an agent inside ChatGPT designed for complex work. It has Codex-technology which it uses as a harness, which means it can do multi-step tasks and take on long-running workflows much better than the standard chat model. It can take a goal, plan out steps, work with connected apps, and do work over several hours. It’s live on ChatGPT Pro, Enterprise, and Edu, with Plus coming soon.
The new Work mode lives directly inside ChatGPT. Source: OpenAI
Why it matters
As for GPT-5.6: the performance-battle between Anthropic and OpenAI remains intense. Anthropic launched Fable 5 just a couple of weeks ago, touting it as the world’s best model, and now OpenAI just launched something very similar at a third of the price. The efficiency aspect is hard to overstate: it lets OpenAI put its flagship inside the regular ChatGPT subscription, whereas Anthropic is restricting Fable 5 to usage-credits (read: super expensive) starting Jul 19.
To be specific: While Sol is included from the ChatGPT Plus plan and up, the maxed-out version (Sol Pro) needs a Pro subscription.
Performance-wise, for anyone who wants the best AI for their daily work without paying extra usage credits, ChatGPT might be the best alternative going forwards.
As for ChatGPT Work: If you’re still only using standard ChatGPT for your work, and not yet using Claude Cowork or an agentic coding tool, then ChatGPT Work is going to feel like a big upgrade. I personally don’t think I’ll be using it much as most of my work happens in AI coding tools (mostly Claude Code) at the command line, and I don’t want to be limited by a UI.
IN PARTNERSHIP WITH NEWSCATCHER
If you're monitoring a portfolio, a supplier network, or any other curated list of companies, CatchAll's Company Watchlist is built for you.
Upload your companies once. Then attach the list to any search job: product recalls, data breaches, or anything else you can describe. CatchAll figures out what events to look for and returns every match it can find. What you get back is a structured dataset of events scoped to your list, with details like what happened, when it was reported, citations, and a relevance score per match.
Run it once, or turn on a monitor and new events are delivered to you as they're reported.
Useful for M&A surveillance, supply chain risk, portfolio monitoring, counterparty tracking, or anything else worth watching across 2B+ web pages.
NEWS NEWS NEWS ❦ NEWS NEWS NEWS
All the small things
Industry moves
Apple is suing OpenAI, saying OpenAI stole its secrets for the AI device it’s building with Jony Ive. OpenAI has hired a stream of Apple hardware people, including its hardware chief (24 years at Apple). Apple says the poaching came with homework: candidates were allegedly told to bring “actual parts” and prototypes to job interviews. OpenAI’s response: “We have no interest in other companies’ trade secrets.” 🫠
Nvidia has lost about $1 trillion in market value in under two months. The pattern is easy to read: OpenAI and Meta are building their own chips, and every lab is chasing cheaper tokens. The market is repricing what happens when your biggest customers start making their own shovels.
The news outlets suing OpenAI say it hid evidence for two years. OpenAI had long claimed it couldn’t search its ChatGPT logs for the copyright case. Turns out it could: one of its own engineers revealed the company had already run exactly those searches. Sanctions are on the table, and the case just shifted from “was this fair use” to “can the court trust anything OpenAI says.”
Models
Grok 4.5 is the first model from the SpaceXAI and Cursor tie-up. It benchmarks alongside Opus 4.8 and GPT-5.5 on coding and agent work, costs $2/$6 per million tokens (less than half of Opus), and it’s temporarily free inside Cursor. The strategy is the bigger story: Cursor owns one of AI’s most valuable workflows, and it can now quietly route your work to a model it co-owns.
Meta started charging for its models. Muse Spark 1.1 is the first one behind the new paid Meta Model API: an agent-focused model with a million-token context window, priced at about a quarter of what the top labs charge. Zuckerberg calls the competition’s margins “very extreme.”
Meta launched Muse Image, then pulled its most controversial feature within days. At launch, anyone could @-mention a public Instagram account and pull that person’s photos into AI-generated images. The backlash was instant, and Meta reversed course: the feature “missed the mark, so it’s no longer available.” What’s left is the image model with 30 new AI effects for Instagram Stories.
New tools & product features
ChatGPT’s new voice mode talks and listens at the same time. GPT-Live replaces turn-based voice with a full-duplex model: a phone call instead of a walkie-talkie. It handles interruptions, hums along while you talk, translates live, and quietly hands hard questions to a bigger model in the background. Paid users get it as the default; free users get a mini version. Haven’t tried it myself yet, though I hear some people complain it’s a little too quick to chat back and easily gets distracted by background noise.
Anthropic published official recipes for cutting your Claude bill roughly in half. Two patterns: let cheap Sonnet 5 do the work and call in Fable 5 only at key decisions, or have Fable make the plan and delegate to parallel Sonnet workers. Anthropic’s own numbers: 90+% of Fable-level quality at around 50% of the cost. Everyone’s anxious about Claude usage limits right now, so Anthropic productizing frugality is smart, and a little telling.
Research
OpenAI checked the industry’s favorite coding benchmark and found a third of it broken. Roughly 30% of SWE-Bench Pro tasks have defects: hidden requirements, tests that reject correct answers, contradictory instructions. Scores had jumped from 23% to 80% in eight months, which now looks less like progress and more like models learning the noise. OpenAI formally retracted its endorsement of the benchmark.
❦
You are a delight.
What's your verdict on today's email?
See which AI use cases are paying off with Context Windows Pro
Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.
I’ve created Context Windows just so you can pick the winners.
🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.