Altman's 5% sacrifice to Washington

Welcome back! OpenAI floated handing the Trump administration 5% of the company, about $43 billion of goodwill, to "clear political obstacles." Benchmarks say the Fable 5 that returned this week is a shadow of the one that left. Tesla capped employee AI spending at $200 a week, with one telling exemption. And Microsoft, days after prepping its July cuts, funded a 6,000-person army that installs AI for other companies.

In today's Generative AI Newsletter:

Sam Altman's plan for getting the government out of OpenAI's way is to bring it inside.

The FT reports OpenAI has discussed giving the US government a 5% stake, about $43 billion at its $852 billion valuation, to "clear political obstacles." 

Altman suggested the number himself, in talks that included Trump, Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent.

The shape he floated is an Alaska-style fund. Every big American lab would put 5% of its equity into a vehicle that pays dividends to the public, the way Alaska pays residents from its oil money. 

Anthropic, Google and Meta would be expected to match, and nobody has said they will.

There's a precedent, and it worked. The government took 10% of Intel last year, and Trump went from attacking Intel's CEO to championing the company.

The talks are early, "conceptual," and might need an act of Congress. Altman has also been talking to Bernie Sanders, who wants the public share closer to 50%.

Last week the government decided who gets GPT-5.6 first. This week OpenAI offered it a seat on the cap table.

Once the Treasury owns a slice of the company it regulates, every AI decision in Washington comes with a conflict of interest built in, and that might be exactly the point.

Anthropic raised Claude API rate limits for all users overnight and simplified its tiers, which no longer depend on API spend.

The newest Sonnet and Haiku models now get 5x higher limits at the highest tier.

The ceiling on how much Claude you can push through the API at once went way up, and developers no longer have to spend their way to higher limits.

Generous timing, because everything else Anthropic said this week points the other way.

Fable 5 came back Wednesday with new safety classifiers, and BridgeMind re-ran the tests.

And it leaves subscription plans after Tuesday, July 7, moving to pay-per-use credits. A Claude Code lead engineer spent the overnight hours reassuring users it comes back "as soon as capacity allows."

Less than an hour later, the 5x announcement dropped. Scarce capacity for the model on your subscription, five times the rate limits for the models that bill by the token.

Tesla spent six months pushing employees to use more AI, complete with internal leaderboards ranking engineers by token consumption. Starting Monday, everyone gets $200 a week unless they use Grok.

Some engineers were burning "thousands of dollars' worth of tokens each week," so spending above the cap now needs sign-off. 

Grok is of course exempt. The favor is worthless because it seems like nobody wants it. Grok is unpopular with Tesla's own engineers, many of whom prefer Claude.

Tesla is the newest name on a list that's becoming a genre. Uber capped engineers at $1,500 a month after finishing its 2026 AI budget in April. 

Meta, Amazon and Walmart have all imposed caps or pushed staff to cheaper models. 

Anyways, budgets of $250 to $2,000 a month per employee are becoming standard. 

Most workers never hit them, and Anthropic's own documentation puts the average Claude Code developer at $150 to $250 a month. 

If they hadn't started asking employees to "tokenmaxx" in the first place, there would be no problem to solve.

Microsoft is putting $2.5 billion into a new unit, Microsoft Frontier, that embeds 6,000 of its people inside customer companies to make AI actually work.

The recruits are solution architects, deployment engineers, trainers and industry strategists, parked with clients for six months to a year at a time. 

In one pilot, that hands-on treatment reportedly cut a supply-chain Copilot rollout from 14 months to 5. Rodrigo Kede Lima, who ran Microsoft's Asia business, leads it.

Palantir built its whole business on engineers who move in with the customer, and now Microsoft, Amazon, Anthropic and OpenAI have all launched some version of it this year.

The honest part came from Judson Althoff, Microsoft's commercial chief, who says the embedded teams will swap customers onto cheaper models when those get the job done. 

Ello is an AI reading coach for kids in kindergarten through third grade. It listens while your kid reads a real book out loud, helps when they get stuck, corrects gently and adapts the books to their exact phonics level. It came out of speech recognition built specifically for children's voices.

Try this yourself:

📸 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.