Last Week in AI: Washington Wants In

Hey there,

I was sick last week, so this covers two weeks. The common thread is the US government showing up in AI news over and over:

• Anthropic brought Fable 5 back. The export controls that forced the suspension were lifted, and Fable returned everywhere on July 1.
• OpenAI previewed GPT-5.6 — three models: Sol, Terra, and Luna. It's a limited preview for selected partners at the government's request. Broader access is supposed to come in the next few weeks.
• OpenAI is reportedly in early talks to give the US government a 5% stake, per FT reporting. They published a public wealth fund proposal earlier this year, so this isn't out of nowhere.

On the stake: most people outside tech haven't seen massive benefit from AI yet. They've tried the free tier and mostly hear about job losses and data centers stealing water (doesn’t matter if it’s true). If regular people owned part of these companies — through something like Trump Accounts — I think the politics change. I'd guess Anthropic looks at something similar. The part I'm not sure about is government involvement in releases becoming normal.

One update from me:

Sonnet 5 and GLM 5.2 aren't good at sales. I ran both through salesevals.com last week and posted the results. They're great for coding at a fraction of the cost of Opus, but they scored poorly on sales coaching — and I'd extend that to anything that requires critically examining the human element, like writing or speaking. Someone replied asking what specifically flunked. Fair question. The answer is longer than a tweet, so I'm turning the failure notes into a blog post on salesevals.

As for next week, that blog post, plus GPT-5.6 through the evals if access opens up. Oh and Fable leaves the subscription plan after the 7th so tokenmaxx away.

Happy 4th 🇺🇸

Drew