The AI race got less shiny this week and way more practical.

Fable 5 and Mythos 5 stayed offline. Microsoft looked at DeepSeek V4 to make Copilot Cowork cheaper. Z.ai pushed open weights closer to frontier coding models.

Then things got weird in a useful way.

Midjourney jumped into full-body ultrasound scans. Cursor started aiming at GitHub. OpenAI may be days away from GPT-5.6.

On today’s Everyday AI, we broke down the real strategy hiding underneath all of it: who gets access, who controls cost, and who owns the workflow.

The week started with a reminder most companies would rather ignore: model access is not guaranteed.

Anthropic’s newest models, Fable 5 and Mythos 5, were still offline after U.S. export controls restricted access outside the U.S. and for foreign nationals inside the U.S. Because Anthropic couldn’t instantly verify every user’s citizenship, it pulled both models globally to stay compliant.

According to Axios, President Trump said he had previously viewed Anthropic, or CEO Dario Amodei, as a national security threat, though he now believed things were improving.

Amazon reportedly flagged concerns. The Pentagon labeled Anthropic a supply chain risk, and Trump did not rule out emergency powers under the Defense Production Act.

What it means

Model access is now a business continuity issue.

If your company depends on one frontier model, build the fallback before something breaks. Approved backup models, vendor risk checks, and swappable workflows are not “nice to have” anymore.

Access was the first pressure point.

Cost was the next.

Microsoft is moving Copilot Cowork to usage-based pricing as it becomes generally available. At the same time, Microsoft is reportedly exploring DeepSeek V4 as a cheaper self-hosted option for the assistant.

The cost gap is the loud part. DeepSeek V4 could cost about 87 cents per million tokens, compared to Anthropic’s $50 per million tokens.

Sheesh.

Any DeepSeek option would run on Azure, stay optional, and include Microsoft safety and compliance filters. Still, the move is spicy because Microsoft works closely with OpenAI and Anthropic, while both have raised concerns about DeepSeek and other Chinese models.

What it means

Token costs are becoming a real operating expense.

Business leaders should stop asking only which model is smartest. Ask which model is good enough, safe enough, and affordable enough for repeated agentic work.

Microsoft’s cost math leads right into the bigger shift: open models are getting harder to dismiss.

Z.ai released GLM-5.2, a 753 billion parameter open-weight model with an unrestricted MIT open source license. It is built for long coding and engineering tasks, with a 1 million token context window for large files, documents, and workflows.

The model is available through Hugging Face, Z.ai’s API, and more than 20 third-party coding environments.

Performance is the part worth watching. GLM-5.2 is challenging top proprietary models on several coding benchmarks and even beating Fable 5, Opus 4.8, and GPT-5.5 on some niche front-end coding tests.

Big catch: it is text only.

What it means

Open models deserve a real test inside your AI stack.

They may not fit multimodal work, but for text-heavy coding and engineering tasks, they could cut costs while keeping performance close enough for practical use.

As open models gained ground, governments were paying attention too.

AI was front and center at the G7 summit in France, where leaders from Anthropic, Google DeepMind, and OpenAI met with G7 leaders and top U.S. officials. The main push: a U.S.-led international AI coalition.

Dario Amodei and Demis Hassabis urged global cooperation around AI standards and risk. Sam Altman pushed for a global forum focused on testing, risk analysis, and international collaboration.

The talks included access to advanced models, chip trade, critical components, cybersecurity, bioterrorism, intelligence risks, and excluding China from certain exchanges.

So yeah, this is not just a product roadmap conversation anymore.

What it means

AI governance is now tied to national security and supply chains.

Companies should expect more rules around model access, chips, data, and vendor risk. AI strategy can’t live only with technical teams anymore.

After the geopolitics, Midjourney delivered the curveball.

The company is moving from AI images into hardware with the Midjourney Scanner, a full-body ultrasound device designed to capture internal images of muscle, fat, bone, and organs.

CEO David Holz said the scanner aims for MRI-like image quality in about 60 seconds. It uses 40 Butterfly ultrasound-on-chip imaging modules and two petaflops of processing power through a partnership with Butterfly Network.

The first planned rollout is not a hospital. Midjourney wants a San Francisco Union Square spa with 10 scanners, a gym, saunas, cold plunges, and hot tub-equipped scanning rooms.

Wild? Yes.

Totally random? Not really, given Holz’s Leap Motion hardware background.

What it means

Midjourney is starting with wellness because body composition maps avoid FDA clearance.

The business question is demand. Repeated scans may make sense to early tech adopters, but price and mainstream interest will decide whether this becomes more than a Silicon Valley experiment.

Midjourney is chasing hardware.

Cursor is chasing the engineering stack.

At its Compile conference, Cursor previewed a from-scratch model with more than 1.5 trillion parameters, trained on more than 100,000 GPUs. The goal is a broader engineering assistant that can plan, test, code, and interact with user interfaces.

Cursor also announced Origin, a Git-native code hosting platform built for AI agents and designed to challenge GitHub.

That matters because Cursor is not just improving the coding surface. With its SpaceX connection, major compute access, and previous Composer 2.5 cost-efficiency story, Cursor is trying to move closer to the center of AI software development.

What it means

Cursor wants to own the AI engineering workflow.

If agents are writing, testing, and merging code, the code hosting layer has to change too. This is about controlling where software work actually happens.

And because one major model week apparently was not enough, OpenAI may be next.

Reports and leaks point to GPT-5.6 arriving soon, possibly with standard, mini, and pro versions launching together.

Some Pro users reportedly saw better technical prompt understanding, though web development issues remained. The rumored context window jumps to 1.5 million tokens, up from 1 million in GPT-5.5.

Developers who believe they tested GPT-5.6 Pro also said it beat Anthropic’s Mythos models on agentic coding tasks.

Add in the possible return of Fable 5 and Google’s expected Gemini 3.5 Pro, and this could turn into a very busy model stretch.

What it means

New model releases should trigger testing, not panic spending.

Compare GPT-5.6, Fable 5, and Gemini 3.5 Pro on real work once available. Use actual workflows, usage caps, and cost tracking. No toy prompts. No guesswork.