Remember when AI was for everyone? That ended this month.

June was supposed to drop frontier models in all our laps, but instead the U.S. government stepped in and quietly started deciding which companies get the good stuff and which ones wait in line.

OpenAI launched its best model ever and handed it to almost no one. Anthropic clawed one model back, but only for a lucky 100 companies. And Google, the company that arguably kicked off this whole era, blew its own deadline while its brightest minds slipped out the door.

Locked-out labs, a half-billion-dollar jobs play, and a China twist that should worry anyone choosing a model.

We untangle all of it on today's episode of Everyday AI.

OpenAI's strongest model ever just launched to a small group of trusted testers only, after the government pushed for restraint and the company quietly cleared its partner list with Washington before release.

There's new naming, too. Think sun, earth, moon: Sol the flagship, Terra the everyday middle, Luna the fast and cheap one.

Sol goes further than the rest, adding a max reasoning mode plus an ultra mode that coordinates sub-agents for heavy coding, biology, and cybersecurity work.

OpenAI claims Sol beats Anthropic's Claude Mythos 5 on coding while burning roughly one-third of the output tokens, though there are no full benchmarks yet, so hold that one loosely.

Safety moved, too. Instead of a separate filter shunting risky prompts off to older versions, the guardrails now live inside the model itself.

What it means: Welp, this is the week truly democratized AI quietly died.

The frontier now goes to the vetted few while everyone else waits on a watered-down version, and since talk of broad access is cheap right now, the only thing worth trusting is what actually reaches your hands.

Anthropic caught a break Friday when the government agreed to let it re-release Mythos 5 to roughly 100 companies and one federal agency, ending a tense two-week standoff with the Trump administration.

Commerce Secretary Howard Lutnick said appropriate safeguards are in place for trusted partners. Fable 5, though, stays benched, which means Anthropic's most advanced public model is still completely dark.

Both got yanked earlier this month after an export-control directive cited national security and cut off access for foreign nationals, a sweep wide enough to include the company's own foreign-national employees.

Here's the juicy part. Talks reportedly moved off CEO Dario Amodei and onto co-founder Tom Brown, after the White House struggled to get anywhere with Dario.

Things have thawed since, and yet the fight is far from over, because Anthropic is still suing the Defense Department over getting branded a supply chain risk.

What it means: Getting your model back as a favor to 100 firms is not a win.

Anthropic is bargaining from a weaker seat than OpenAI, juggling a partial re-release, a benched model, and a live lawsuit all at the same time, which is exactly why the bigger lesson lands so hard: the government holds the on-switch now, so rethink what you're building on top of.

Google promised Gemini 3.5 Pro in June at I/O, and here we are in the final hours of the month with nothing to show.

Government pressure is part of it. The bigger story, per The Information and others, sits inside the building.

To close the gap with Anthropic, Google is making its short-term coding strike team permanent, and that new mid-training stage is meant to teach the model far more than syntax: bigger engineering tasks, tool use, and reading across multiple files at once.

The why stings. Internal estimates reportedly put Google engineers near 50 percent AI assistance against nearly 100 percent at Anthropic.

Then came the exits, with Jonas Adler, Alexander Pritzel, John Jumper, and Noam Shazeer all reportedly bolting for rivals.

Sheesh. Some of the biggest names in the game, gone.

What it means: A code red by a friendlier name is still a code red.

Google helped invent this stuff and is now chasing it on its own turf, aiming squarely at long-running engineering agents instead of autocomplete, but hitting that goal while your sharpest minds keep defecting to the very labs you're trying to catch is a brutal way to run a race.

A new nonprofit called RAISE US has already raised $500 million toward a $1 billion goal aimed at helping states brace for AI disruption before it fully arrives.

Former Commerce Secretary Gina Raimondo leads it, and her framing is blunt. America has a tech strategy for the AI race but no people strategy for workers.

The backers say plenty, because OpenAI Foundation, Anthropic, Amazon, and Microsoft are all anchoring it.

The plan itself is a national platform that coaches governors on prepping local workforces through practical policy and training.

Four states are already in, split evenly red and blue, with Arkansas working alongside Governor Sarah Huckabee Sanders on a career navigation platform that ties job seekers to employer-backed paths, and Maryland teaming with Governor Wes Moore on service-year roles across healthcare and education.

What it means: The same firms automating the work are now funding the airbag, and that tension is the whole story.

We've said since day one that AI rollout is change management more than tech installation, and this is that exact idea dressed in a half-billion-dollar suit, which leaves one real test: whether workers actually land better jobs, or it just buys a season of good headlines.

Per CNBC, Anthropic told the Senate Banking Committee this month that Alibaba pulled off what it calls the largest known distillation attack on its models.

The scale is staggering: roughly 28 million exchanges through about 25,000 fraudulent accounts between April and June. Anthropic says operators tied to Alibaba and its Qwen AI lab ran the whole thing to copy how its models behave.

Quick refresher. Distillation is when a weaker model learns from a stronger one's outputs, and Anthropic is framing this particular version as flat-out theft rather than fair play.

The timing is loaded, landing two months after the White House pledged to help labs catch industrial-scale distillation, with Anthropic arguing Alibaba ignored the warning entirely.

So now it folds straight into the larger U.S.-China tech brawl.

What it means: This smashes right into the access story.

Lock the frontier to a chosen few here, while Chinese open-weight models like GLM-5.2 from Z.ai keep climbing into that same tier, and the calculus around what your company runs shifts fast, because next year's model choice could hinge on who's allowed to sell it and who stands accused of stealing it.

OpenAI and Broadcom unveiled Jalapeño, a custom chip built to run large language models cheaper, faster, and less power-hungry at scale, with deployment expected later this year.

That efficiency matters more than it sounds, because energy and compute costs are fast becoming the real ceiling on how widely these labs can serve frontier AI at all.

Jalapeño is built for inference ragther than training, so it's tuned for the endless back-and-forth of answering users. Nvidia still owns training, and this chip goes after everything else: speed, memory, networking, and wasted compute.

The companies say it beats today's top hardware, though they've shown no benchmarks. The real shocker is pace.

Early samples are already running in the lab, and the thing went from blank page to working silicon in just nine months.

What it means: Own the chip, own your costs.

When you're serving close to a billion users a week, every sliver of a cent shaved off inference quietly compounds into serious money, and a nine-month build shows just how fast this race is sprinting, because the more of its own stack OpenAI controls, the less it leans on Nvidia and the more room it has to set its own prices.