Claude's back everywhere, Washington wants a seat, and a food delivery app drops a free model

First issue. Here is the deal. Every Tuesday I round up what actually happened in AI the week before and what it means if you build with it. I run my own projects on these tools every day. So this is filtered through one question: does this change what I do Monday morning? Not what sounds exciting in a headline.

Here is what mattered last week.

Claude Fable 5 came back for everyone

The Commerce Department lifted the export controls on June 30. Anthropic turned Fable 5 back on worldwide on July 1 through the 7th. Claude.ai, the API, Claude Code, all of it. If you dropped down to an older model during the restriction, you can move your jobs back up. Anthropic also shipped Sonnet 5 the same week and called it their most agentic Sonnet yet. Worth a test if you run high-volume automated work where Fable pricing adds up.

What it means for builders: model access is now a supply chain risk like anything else. This was the first time a frontier model got pulled and switched back on by a government decision. It will not be the last, look at Google…If your business runs on scheduled agents like mine does, know exactly which model each job uses and what your fallback is.

Washington wants to write the rules of release

The Financial Times reported the White House is deep in talks with OpenAI, Google, and Anthropic on voluntary standards for how frontier models get released. An announcement could land this week. At the same time OpenAI is reportedly offering the government a 5% stake ahead of a planned September IPO.

What it means: “voluntary standards” is how an industry buys time before real regulation shows up. Expect release schedules to get slower and more predictable. If you build on top of these models, predictable beats fast. This is good for you.

The free model surprise of the week

Meituan, the biggest food delivery platform in China, released LongCat-2.0 under an MIT license. Think about that. A logistics company shipped a serious open model under the most permissive license there is. That tells you where open weights are heading. They are becoming a side effect of running at scale, not a product someone sells.

Two more worth watching. xAI’s Grok 4.5 went into private beta inside SpaceX and Tesla. OpenAI previewed a GPT-5.6 line called Sol, Terra, and Luna, starting with trusted partners. Nothing you can use today. But the naming tells you both companies are moving to families of specialized models instead of one flagship.

What I did about any of it

Almost nothing. That is the point. My calculator site shipped its Monday tool and its Thursday articles right through all of it, because the pipeline pins its models and checks its own output against fixed rules. Build it so the big news weeks are boring weeks.

More next Tuesday. If someone forwarded you this, subscribe at blog.jaredhebb.com/newsletter.

Jared