Welcome back! The three new OpenAI models, better than last gen yet priced cheaper, drop tomorrow after USA greenlit the release. Meta's superintelligence lab shipped its first image model straight into WhatsApp, and Musk's first Cursor model could land any day. A lot of things got out of the lab this week.
In today's Generative AI Newsletter:
OpenAI: What happens Thursday to the models the government held back?
Claude: Why did they take so long to catch up to Codex?
Meta: Why is Zuckerberg's superintelligence lab making Instagram filters?
Cursor: Whose model lands inside Cursor this week?
Two weeks ago you needed to be a government-vetted partner to touch GPT-5.6. Starting Thursday you need a login.
OpenAI confirmed Sol, Terra and Luna launch publicly this Thursday across ChatGPT, Codex and the API. The June preview ran restricted because the US government asked for an early look, and Axios reports the Commerce Department has now lifted the hold. It lasted 13 days.
The naming is new. The number is the generation, and the names are permanent tiers:
Sol: the flagship and OpenAI's strongest model yet, $5 in and $30 out per million tokens, with a new ultra mode that splits work across subagents.
Terra: matches GPT-5.5 at half the price by OpenAI's own benchmarks, $2.50 in and $15 out.
Luna: the fast, cheap tier at $1 in and $6 out.
Washington's interest was Sol's hacking skill.
It goes toe to toe with Anthropic's Mythos Preview on exploit benchmarks using a third of the output tokens. That's the capability the government wanted to see first.
The same government banned Anthropic's Fable 5 for three weeks in June. OpenAI cooperated and walked out with a launch date and a cheaper lineup. Frontier releases now include a Washington step, and this is what passing it looks like.
Special highlight from our network
Most people use 5% of Claude. Master the rest in one weekend.
Most professionals open Claude, ask a question, and close the tab. That's about 5% of what Claude can actually do in 2026.
The other 95% is where the interesting work is happening. Skills that give Claude persistent expertise. Connectors that plug it into Notion, files, and your desktop. Cowork for full document and project workflows. Vibe coding for building apps without being a developer.
Outskill's 2-Day Claude AI Mastery Workshop is a live end-to-end walkthrough of all of it, condensed from 800+ hours of research into a focused 16-hour curriculum. Plus 10+ AI tools and workflows that pair with Claude.
Saturday and Sunday, 10 AM to 7 PM EST. Free for the next 48 hours.
The best people to learn AI from are the ones who actually use it better than others. When you have a team full of these experts, you want to share such productive knowledge with everyone.
Explore the options by clicking on the title to find out what they put out there and what you can achieve by just watching a couple resources.
Claude Cowork, Anthropic's hand-it-a-task agent, is rolling out to mobile and the web.
Work now keeps running in the background with no device online at all, so closing the laptop no longer stops it.
Set Monday's client prep for 6 am and Claude works the email threads and transcripts, builds the briefing and leaves the follow-up drafted but unsent. You review it over coffee.
One early user, a Ramp customer success manager, started a dashboard on his laptop and picked the session up on his phone at baggage claim. "It just held the thread."
Beta hits Max plans first over the next several weeks. Desktop stays the only place Claude can touch your local files and browser, and doubled usage limits run through August 5.
Anthropic says more than 90% of Cowork usage has nothing to do with code. This launch is Anthropic admitting who its agent works for.
The lab Mark Zuckerberg built to chase superintelligence shipped its first image model yesterday. It makes pictures in WhatsApp.
Muse Image comes from Meta Superintelligence Labs, the unit run by Alexandr Wang, the Scale AI founder Meta hired to lead its AI push.
It's free in the Meta AI app, US Instagram Stories and WhatsApp in some countries, and it powers more than 30 new Instagram effects.
It works like an agent. It searches the web for references, writes code when a detail has to be exact, a QR code that scans, and critiques its own drafts before you see them.
It ranks No. 2 on Arena, behind only OpenAI and ahead of Google's Nano Banana 2.
The fight shipped with it. Muse can pull photos from any public Instagram account into your generations, and the setting is on by default. Account owners have to find the toggle to opt out.
SpaceXAI and Cursor plan to launch their first jointly built model as soon as this week, per a staff memo reported by The Information. Cursor declined to comment.
It's the model SpaceX promised when it agreed to buy Cursor for $60 billion in stock last month, days after going public. It ships inside both Cursor and Grok Build, SpaceXAI's own developer product.
Cursor's Composer models got near the frontier cheaply by scaling reinforcement learning, and the company said in April it had been bottlenecked by compute.
Colossus, the supercomputer that came with the xAI acquisition, removes that ceiling.
Cursor got big as neutral ground, the editor where you pick any lab's model. Its owner now stocks a house model in that same picker. Watch the default dropdown over the next month. It belongs to a competitor of everything else in it.
Boris FX makes the effects and cleanup tools behind work at Netflix, the NFL and ILM, and it opened July by buying iZotope, the RX audio repair house. Its pitch now covers picture and sound under one roof.
Try this yourself:
Go to their website and install the Boris FX Hub. One installer covers every product, free trials included.
Got a noisy podcast track? Run CrumplePop's AudioDenoise or EchoRemover on it and compare the before and after.
Retouch photos? Open Optics from Photoshop, Lightroom or Apple Photos and let Face ML auto-mask eyes, lips and skin instead of drawing masks by hand.
Editing video? Drop Continuum's AI masking on a clip in Premiere or Resolve to isolate a subject without keyframing.
Who it's for: video editors, podcasters and photographers who lose hours to manual masking and audio cleanup.
Mid-market companies get their own AI stack: Accenture's two-week-old Edge unit teamed up with Google Cloud on agentic AI for companies making $300 million to $3 billion a year, a market Accenture sizes at $240 billion.
The AI giants are paying startups to pick them: The WSJ reports OpenAI and Anthropic are handing startups free compute to win their business, with one founder stacking offers worth more than $3 million and Anthropic's standard Y Combinator offer jumping from $30,000 to $500,000.
Microsoft started swapping its partners out of Office: Bloomberg reports tens of thousands of prompts a week in Excel and Outlook now run on Microsoft's own MAI models, and AI chief Mustafa Suleyman says the goal is to "reduce and ultimately eliminate" the Anthropic bill.
China may keep its best AI at home: Reuters reports Beijing met with Alibaba, ByteDance and Z.ai about restricting overseas access to China's top models, including unreleased ones, and floated making model leaks a national security offense.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
