DeepSeek ships a 1.6T open model, China builds its own AI bloc, and Google reportedly hits reset

Every Tuesday I round up what actually happened in AI the week before, filtered through one question: does it change what I do Monday morning. Not what sounds big in a headline.

Two stories this week were real and shipped. One was a rumor loud enough that people are already planning around it. Here is what mattered.

DeepSeek shipped V4, and it prices tokens by the clock

DeepSeek moved V4 from preview to its official mid-July release. Two open-weight models under the MIT license, no strings. The big one, V4-Pro, is 1.6 trillion parameters with only 49 billion active on any task, so it runs far lighter than the headline number suggests. A smaller Flash version sits under it. Both take a million tokens of context, and both flip between a fast mode and a slower thinking mode.

The part worth your attention is the pricing. DeepSeek is charging different rates for peak and off peak hours. Run your jobs when the servers are quiet and you pay less. Run them at the busy hour and you pay more.

What it means for builders: this is time-of-use electricity pricing, applied to AI. If your work is batch and not urgent, drafting, summarizing, overnight processing, you can schedule it into the cheap window and cut the bill without touching a line of code. MIT weights mean you can also run it yourself if the math favors that. The hour of the day just became a cost lever you control.

China stood up its own AI governance body

On July 17 in Shanghai, Xi Jinping opened the World AI Conference in person for the first time. He called AI development “a symphony of global cooperation,” not a solo by any single country, and pushed back on what he called the overstretching of national security concerns. Then came the substance. 29 countries signed on to a new World AI Cooperation Organization, an intergovernmental body headquartered in Shanghai. China also promised 5,000 AI training slots for developing countries and open access to a Chinese weather-prediction system for 30 nations.

Set that next to what I wrote about a few weeks back, the White House working on voluntary release standards with OpenAI, Google, and Anthropic. Two blocs, two playbooks. One leans on a handful of American labs. The other pitches open access to everyone outside the export-control fence.

What it means for builders: the ground under model access is splitting into camps, and where you can legally run which model is turning into a real question instead of a footnote. If your product ships internationally, the origin and license of your models is a business decision now, not only a technical one. Watch which models arrive with territory carve-outs and which, like DeepSeek this week, arrive with none.

Google reportedly tore its next model down and rebuilt it

Here is the one to hold loosely. Multiple outlets reported Google scrapped the base model for Gemini 3.5 Pro and restarted pretraining after engineers found it failing on recursive tool calls and image generation, pushing the launch to mid-July. The talked-about specs, a two million token context window and a deeper reasoning mode, are leaks. As of this writing Google had not published a model card, pricing, or a release note. Treat all of it as unconfirmed.

What it means for builders: two lessons hold even if the details turn out wrong. First, do not rebuild your stack around a model that has no official release post. Leaks slip. Second, if the rebuild story is true, it is a tell. The frontier labs are now willing to eat a public delay to fix reliability on tool calling, because agents that call tools are where the money is going. That is the same thing I care about on my own site. A model that writes code is worth nothing if the button it wires up does not fire.

What I did about any of it

The cheap-and-open trend is not abstract for me. My whole pipeline runs on models handing back code that I check. This week it shipped four calculators, and I put a short electrical field guide on Kindle, assembled from work the tools had already done. It also picked what to build next by reading real search demand instead of my gut, and it wired up an email that sends a free formula sheet to every new subscriber on its own.

None of that needed a frontier model. It needed a cheap one that does the job and a set of checks that catch it when it is wrong. DeepSeek pricing tokens by the hour is a nice lever. It is not the thing that lets me sleep at night. The test that fails the build when the math is off does that. Better and cheaper models raise the ceiling on what one person can ship. They still do not press the button and look at what actually went out. That part is the job.

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

Jared