Writing — The Week in AI

The Week in AI: The Frontier Split in Two

November 13, 20266 min readBy Milan Cheeks

This week in AI, the frontier split into planners and workers, decisions became their own product, and agents moved onto factory floors.

The Week in AI: The Frontier Split in Two

Every week in AI brings a launch. This week brought a split, and I think the split matters more than any single launch.

For a few years the story was simple. One big model tried to do everything, at one high price. This week the biggest labs moved the other way. They separated the thinking from the doing, priced each part on its own, and started selling the routing between them.

A top model, and a cheap worker beside it

On November 9, Anthropic released Claude Fable 5.2, alongside Claude Mythos 5.2, a restricted sibling for vetted organisations in cybersecurity and life sciences. Fable costs 8 dollars per million input tokens and 40 dollars per million output tokens. On the Artificial Analysis Intelligence Index it scored 65, ahead of Google Gemini 4 Carbon at 60 and Anthropic own Claude Opus 5.5 at 59.

The same day, Meta released Muse Ember, its cheapest model yet, at 0.10 dollars per million input tokens. Meta built it for the repetitive steps of agent work, such as tool calls, data extraction, and reading long documents, while a larger model plans.

The shape is plain. The frontier is no longer one model. It is a team, with a planner at the top and workers below it, and the price gap is now wide enough to design around. Sending every step to a flagship model means paying flagship prices for chores.

The question is no longer which model you use. It is which model you use for each step, and who decides.

Google made the worker the default

On November 12, Google made Gemini 4 Hydrogen generally available and the default model for high volume work in its tools. It costs 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, across a context window of one million tokens. Google positions it for coding agents and everyday tasks, and its Antigravity platform added an Orchestrator that moves work between models mid task.

A day earlier, Google launched the Gemini Decisions API. Its model, Gemini Decision 1, does not write text at all. It scores bounded choices, such as yes or no questions and rubric scores, over text, images, audio, video, and documents. Google says one request can carry up to 1,000 questions over the same input, at 0.10 dollars per million input tokens with no output charge, and response times it reports at 150 to 250 milliseconds.

That launch is the one I keep thinking about. Much of what agents do is not writing. It is deciding. Which tool to call. Whether a document passes a rule. Whether a person needs to look. Google turned that step into its own cheap product, and a step with its own product and price has become infrastructure.

Agents moved onto your servers

On November 12, Anthropic opened a private beta that lets companies run Claude Managed Agents on their own cloud or on their own premises, including the agent server, credentials, sandboxes, and session history. The weights stay with Anthropic. What moves is the control layer, the part that holds your data, your tools, and the record of what the agent did.

That is the real enterprise story. Companies were never only asking for a smarter model. They were asking where the agent runs and who keeps the record.

Robots got the same split

On November 10, Google DeepMind released Gemini Robotics 2.5 to robotics partners. In a pilot on live Mercedes Benz production lines in Hungary and Germany, Apollo 2 humanoid robots running the model sorted metal brackets into trays in vehicle order, with a 96 percent success rate reported by Google.

On November 11, Runway opened the weights of Praxis-1 under the Apache 2.0 licence, along with its fine tuning code. Praxis-1 is a 27 billion parameter robotics model that predicts video and actions together, and Runway reports a 46.2 percent success rate across 1,200 benchmark episodes in RoboLab.

One story is a closed model proving itself on a factory floor. The other is an open model any team can inspect and adapt. In physical work, that choice matters, because a wrong decision by a robot often cannot be undone.

Security opened up, and oversight got harder

On November 12, OpenAI released Codex Security as open source software, a command line tool and a TypeScript SDK any developer can pull from GitHub. It reviews code for security issues, validates what it finds, and generates candidate patches. Scans can resume after an interruption and run under a cost budget set in advance.

On November 10, OpenAI also published research on a quieter problem: how pretraining, architecture choices, and reinforcement learning affect whether a frontier model reasoning can be monitored through the trace it leaves. In plain terms, stronger models can solve more tasks without leaving a readable trace. Watching the words an agent produces is not the same as watching what it does.

I build platforms where agents create real things for real people, including Manifest Anything, where players describe a game and the system builds it. That research speaks directly to builders like me. Watch the tool calls, the actions, and the outcomes. Do not build a safety story on a transcript.

What I take from this week

Three lessons stay with me. First, design for a team of models, not a single model, and put effort into the routing between them. Second, treat decisions as their own step, with their own cost and checks. Third, the layer that decides where agents run and who keeps their records is becoming as important as the models themselves.

I build in public, so my own numbers stay open to check. My Instagram audience at @milandadon stands at 77,289 followers, and my Threads posts drew 980,205 views across the trailing 30 days. Proof beats promises for a founder too.

Next week will bring more launches. I will ask a new question. Not just what shipped, but which layer it shipped in, and what that layer now costs.

If your conference or team is working through what these shifts mean in practice, I speak on building with AI in the open. Book me for a keynote here.

← All writing Book Milan →
The Newsletter

The frontier, in your inbox.

One email when a new essay drops — the Kardashev roadmap, AI-native building, and coexistence, written as it happens. No noise, unsubscribe anytime.

You're on the list.