AI Actually
Issue No. 23 · Wednesday, July 1, 2026
This week: Anthropic’s most dangerous model came back from the dead, OpenAI released a model so powerful the government made it invite-only, Meta figured out how to turn your thoughts into a keyboard, and Google told Meta — politely, corporately — that there simply wasn’t enough computer to go around.
The model that got sent to the principal’s office is back
Three weeks ago, Anthropic launched Claude Fable 5 and Claude Mythos 5, its most capable public models yet. Three days later, the government shut both of them off — everywhere, for everyone, worldwide — over an export-control order citing “national security.” The trigger, per Anthropic, was a jailbreak that let outside researchers coax Fable into identifying a handful of already-known software flaws. Anthropic argued plenty of other models, including its own weaker ones, could do the same thing. The government disagreed, at least for 19 days.
On Tuesday night, the Department of Commerce lifted the controls. Fable 5 came back online Wednesday, with a new safety filter Anthropic says blocks the reported bypass more than 99% of the time. Mythos 5 access is also widening for approved organizations. In the middle of all this, Anthropic also quietly shipped Claude Sonnet 5, a cheaper, more agentic mid-tier model — timing so awkward that even the people covering it called it “arriving in Fable’s shadow.” Sonnet 5 is a real upgrade over its predecessor. It’s just not the model anyone was waiting for.
Why it matters. A private company built something, called it dangerous, and a federal agency believed it enough to reach for an off switch — then changed its mind three weeks later without ever fully explaining why. Nothing about the underlying model changed in that window except a patch. What changed was the negotiation.
Meta wants to skip the keyboard entirely
Meta unveiled Brain2Qwerty v2, an AI system that reads brain activity and turns it into typed sentences — no surgical implant required. Volunteers wear a magnetoencephalography (MEG) helmet, type normally, and the model learns to map their brain signals directly to words. Across nine participants who each logged about 10 hours in the machine, the system averaged 61% word accuracy; its best subject hit 78%, with more than half of their sentences landing with one wrong word or fewer. That’s up from roughly 8% for earlier non-invasive attempts. Meta has open-sourced the training code so other labs can build on it.
It’s not mind-reading — participants were actively trying to type memorized sentences, and the hardware is a room-sized scanner, not a headband you’ll own. But the gap between “brain surgery required” and “put on this hat” just got a lot smaller.
Why it matters. This is aimed at people who’ve lost the ability to speak or type after a stroke or injury — a real, unglamorous use of AI that isn’t a chatbot. It’s also a preview of a much stranger conversation later: once a machine can read intended words out of your skull with no incision, “private thought” becomes a slightly more negotiable concept than it used to be.
OpenAI built its most powerful model and then mostly hid it
OpenAI introduced GPT-5.6, a three-tier family — Sol (flagship), Terra (a cheaper all-rounder), and Luna (fast and cheap) — with real gains in coding, cybersecurity, and long, multi-step tasks. Sol reportedly beats Anthropic’s Mythos 5 on several agent benchmarks. But almost nobody can use it yet: at the U.S. government’s request, OpenAI is limiting the initial rollout to about 20 pre-approved partner organizations, with a broader release promised “in the coming weeks.”
This isn’t a coincidence. Back in June, we told you the White House was pressuring OpenAI to slow-walk this exact release over safety concerns. Turns out “slow-walk” meant “launch it, but only for people we’ve already met.”
Why it matters. This is the second time in three weeks a frontier lab has shipped its most capable model directly into a government-shaped bottleneck. It’s starting to look less like a one-off and more like the new normal for how the most powerful AI models reach the public: launched, then rationed.
Google told Meta it couldn’t have what it paid for
Back in March, Google informed Meta that it couldn’t supply the full amount of Gemini computing capacity Meta wanted to buy — despite Meta being one of Google’s biggest cloud customers. The shortfall reportedly delayed some of Meta’s internal AI projects and pushed the company to tell staff to use their AI “tokens” more sparingly. Google, meanwhile, is so stretched that it’s now paying Elon Musk’s SpaceX $920 million a month for bridge capacity. Google Cloud’s backlog of paid-for-but-undelivered work nearly doubled last quarter, to roughly $460 billion.
Meta’s response has been to lean harder on Muse Spark, its own in-house model, so it depends less on a company that also happens to be a rival.
Why it matters. Every AI headline this year has been about what these models can do. This one’s about what they can’t get: literal physical computer chips and electricity. Even Google — a company that owns actual data centers — is turning away customers with unlimited budgets. That’s not a software problem you patch. That’s a supply chain, and supply chains take years.
Safe to ignore this week
Devin Fusion & DeepSeek DSpark — coding-agent tooling updates, developer-only interest.
AI Engineer World’s Fair dispatches — “software factories,” “loops,” and other conference jargon best left at the conference.
Claude Science + Anthropic’s drug discovery program — genuinely interesting if you’re a scientist; the rest of us can wait for a cure before we care how it was built.
Nano Banana 2 Lite / Gemini Omni Flash — Google’s cheaper image and video models. Fine tools, not news.
“Is AI killing entry-level jobs?” — real question, but only one outlet ran it this week. We’ll come back when there’s more than a vibe.