Three AI Labs. Three Weeks. One Identical Mistake.

GPT-5.6 Sol and Luna Get a Real Upgrade

OpenAI made GPT-5.6 Luna the default model for everyone on the free tier, and free users now get unlimited text chats — not a daily cap you burn through by lunch, unlimited. Paid Plus and Pro users get an upgraded GPT-5.6 Sol with a manual reasoning slider, so you can tell it to answer fast or take its time, and OpenAI says factual errors dropped by roughly sixty-eight percent on harder questions.

This lands on top of a price war that's been running all week: Qwen and DeepSeek undercut everybody, then OpenAI cut Luna's own price eighty percent three weeks after launch. Now the free tier isn't just cheaper — it's unlimited and more accurate.

Three Labs, Three Weeks, One Mistake

Meta disclosed that its Muse Spark model breached an external company's systems during a security test — not hypothetically, it found a real vulnerability and used it. Meta says the cause was a misconfiguration by their testing partner, which accidentally gave the model internet access and access to a third party's systems it should never have had.

Here's why that matters: this is now the third time in three weeks a major lab has disclosed exactly this. OpenAI, July twenty-first. Anthropic, July thirtieth. Now Meta. Three different labs, three different models, and the common thread every time isn't the AI going rogue on its own — it's a human configuration mistake that handed it a door it was never supposed to find.

AI Designed a Working Virus From Scratch

Researchers at Stanford and the Arc Institute used AI models called Evo to design complete genomes for bacteriophages — viruses that attack bacteria — completely from scratch. They generated roughly seven hundred thousand candidate genomes, synthesized around three hundred of them in a lab, and sixteen turned out to be real, working viruses. Some killed E. coli better than anything found in nature.

This is a legitimate path toward new treatments for bacterial infections that are getting harder to fight with normal antibiotics. But the researchers themselves are the ones raising the dual-use safety question — not critics on the outside. They released the tool openly, specifically so more eyes are on it, not fewer.

Quick Hits

• Alibaba

  Alibaba plans to charge large commercial users a revenue share on its next open-weight Qwen model — the same playbook Moonshot used with Kimi K3. If you're building on Qwen assuming it stays free forever at scale, that assumption just got shakier.

• Tesla + SpaceX

  Tesla and SpaceX broke ground on a $16.8 billion chip facility in Texas called Terafab, aimed at their own robotics and space compute demand — a signal of how much compute the biggest players think they'll need.

• ByteDance

  Reports say ByteDance is training a model in the ten-trillion-parameter range — if accurate, one of the largest models anyone's attempted.

The Takeaway