AI completed a BMW. What's it completing at your company?

In this edition:

• This week: ChatGPT Work ships to 1 billion users, AlphaEvolve goes generally available, and Figure robots complete their first BMW production run

• Under the radar: Fable 5 returned with terms that give the US government standing early access to Anthropic's future unreleased models

• What's on the calendar: Fidji Simo steps back from OpenAI, Cursor building a work agent, and AlphaEvolve now available on Google Cloud

THE WEEK IN AI

THE WEEK IN ONE SENTENCE

OpenAI bet that distribution beats model performance by shipping a work agent to 1 billion ChatGPT users, Google made an autonomous optimization agent generally available, and Figure AI put humanoid robots on a BMW production line. These are three deployments that finished work, not assisted with it.

THREE SIGNALS

01 • Agents

OpenAI turned ChatGPT into a work agent

On July 9, OpenAI introduced ChatGPT Work, powered by Codex and GPT-5.6. The announcement describes it as an agent that "can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work."

The model itself, GPT-5.6 Sol, launched alongside it. Sol costs $30 per million tokens (40% less than Fable) and beats Fable on TerminalBench 2.1. OpenAI also shipped "ultra mode," which coordinates multiple agents in parallel on demanding tasks.

The interesting move isn't the benchmark result. It's the interface decision. OpenAI has roughly 1 billion weekly active users on ChatGPT. Embedding a work agent there, rather than shipping it API-first the way Anthropic did with Fable, means they're betting that distribution beats raw model performance. Fable is the better-reviewed model in developer circles. ChatGPT Work is where the actual business users already are.

For an operator, that distinction matters. The question isn't which model scores better on a benchmark. It's the system your team will actually open when they have a real task to complete.

TEAM READINESS AUDIT

Get your team’s AI readiness score. Leave with an AI readiness brief.

A 15-minute assessment that turns your answers into a real decision artifact: your readiness level, your six-axis shape, where you're strongest, where you're constrained, and what not to build yet.

02 • Models

Google made AlphaEvolve generally available

The same day, Google Cloud announced that AlphaEvolve is now generally available. Co-developed with DeepMind, AlphaEvolve is a Gemini-powered evolutionary agent that rewrites and optimizes code to solve algorithmic bottlenecks. It discovers optimizations that human engineers miss, not by reasoning through the problem, but by generating, testing, and iterating candidate solutions autonomously.

Google has described AlphaEvolve as having already found improvements in production systems internally, including a scheduling optimization that recovers roughly 1% of Google's total compute capacity. At Google's scale, 1% is not a small number.

This is different from a code assistant that completes lines. AlphaEvolve takes a performance problem, defines the solution space, and searches it systematically. It runs for hours. When it finishes, it returns working code.

The operator angle is specific: if your team has known bottlenecks (a slow query, a costly batch job, a scheduling problem), this is the category of tool worth evaluating now, before your competitors do.

03 • Robotics

Figure AI deployed humanoid robots at BMW's Spartanburg plant

Figure CEO Brett Adcock posted on June 30 that F.03 had arrived at BMW's Spartanburg factory. In a follow-up, he described the commercial deployment: six months of F.03 robots working on BMW's body shop assembly line, helping build the X3. He bought the first four cars off the line himself.

This is the first commercial humanoid robot deployment at a major automaker, not a pilot but a production line with ongoing logistics work. The physical AI story is often told in the future tense. This one is past tense.

The reason it matters beyond manufacturing: physical AI deployments at scale generate interaction data that feeds back into training. Every robot collecting manipulation data on a BMW line makes the next generation of robots faster to train. This is the same compounding loop that made language models improve so quickly, except the inputs are physical rather than textual, and the collection rate is limited by how many robots you can actually place and operate.

UNDER THE RADAR

Fable 5 returned globally on July 1 after a 19-day government-mandated shutdown. Most coverage focused on the return. The more interesting story is the terms.

Anthropic agreed to restore Fable 5 with new classifiers that block certain cybersecurity tasks and route "routine tasks like coding and debugging" back to the older Opus 4.8 model. Anthropic also agreed to give government partners early access to future unreleased models.

That last piece got very little attention. It means the U.S. government now has a standing arrangement for preview access to Anthropic's most capable systems before they ship publicly. There are legitimate arguments for why this makes sense: safety review, responsible deployment, classified use cases. There are also real questions about what "early access" means in practice and whether it creates a two-tier capability environment.

The Open Frontier summit, held earlier this week and attended by leaders from Stanford, MIT, NVIDIA, Hugging Face, and 12 other institutions, spent a day discussing how to pull frontier AI research back into the open. The Fable shutdown was the precipitating event. "Open source as business continuity" shifted from a philosophical stance to an operational argument.

If you're procuring AI systems and your vendor evaluation doesn't include a question about what happens if the model gets taken offline, it probably should.

QUOTE OF THE WEEK

"The product layer now matters as much as raw benchmark leadership because the interface decides whether a model can retain context, use tools, revise its own output, and return a useful deliverable."

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

WHAT’S ON THE CALENDAR

• Fidji Simo is stepping back from her full-time role at OpenAI, transitioning to part-time advisor after a medical leave. She led the consumer product side. Who fills that role and what it signals about OpenAI's internal balance between product and research is worth watching over the next quarter.

• Cursor is reportedly building a general-purpose AI work agent to compete with Claude Cowork and ChatGPT Work. The work-agent category is now formally contested by at least four significant players.

• AlphaEvolve is available to Google Cloud enterprise customers today. If you have a performance-bound engineering problem and you're on Google Cloud, this is worth a conversation with your account team.

P.S. I've been working with a handful of operators over the last few weeks on one specific question: how do you evaluate AI systems when every lab is claiming best-in-class? If your company is mid-decision on which platform to build on, hit reply. I'll share what I've been seeing actually work.

Have a good weekend,
Haroon