Lead Claude Code developers leave Anthropic as the product tops $2.5 billion in revenue

Two prominent engineers have left the Claude Code team. On October 24, Thariq Shihipar, one of the most public faces of the project, announced his departure from Anthropic. Anthony Morris, who built the agent's desktop version, quit right after him. The exits come as Claude Code's annual recurring revenue has topped $2.5 billion, according to the company itself. For a product that is only a few years old by industry standards, that is a lot.

Who left and what they leave behind

Shihipar came to Anthropic from a research background: he previously worked at MIT Media Lab and is also a Y Combinator alumnus. He spent just over a year on the project. His departure post drew millions of views. In it he passed no sweeping judgments and instead praised the team that remains. He also suggested following Claude Code updates through Addy Osmani and Lydia Hallie. Shihipar himself plans to take a break.

Morris worked on the agent's desktop version, the layer of the product that takes a coding agent out of the terminal and puts it in front of a wider audience. Losing the person behind that effort is noticeable, but according to reports, product work will continue without pauses.

Two departures in a row at the height of commercial success look like a classic pattern for fast-growing teams. The people who built the first versions of a product gain strong positions in the market and burn out from the pace. Meanwhile, Claude Code remains one of Anthropic's main sources of revenue, so retaining key engineers matters more now than it did before.

What else happened in AI today

Alongside the personnel news, several other stories broke that show where the industry is heading:

• Sakana AI. The Tokyo startup, known for its Fugu system that combined off-the-shelf neural networks, is changing course and taking on training its own model, working title Sakana Sawara. Building on other people's weights meant dependence, and Japan needs technological sovereignty without being tied to US and Chinese corporations. A compact version is promised for spring 2027, after which the model will be scaled up to hundreds of billions of parameters. Target areas are national security, finance, vulnerability discovery and cyberattack analysis, and defense. Sakana is already testing its AI analyst under a contract with Japan's Ministry of Defense.
• Nvidia. Jensen Huang believes software will soon become the most important part of the company's business. A comprehensive license for deploying agents on enterprise AI factories will cost roughly $2000 per accelerator per year. If the subscription covers even five million chips, it would bring in up to $10 billion in steady revenue. How the new rate will coexist with the existing AI Enterprise subscription at $4500 remains unclear.
• xAI and Grok Imagine. In July, Elon Musk announced a full-length, historically accurate screen adaptation of Homer's Odyssey. Instead of a feature film, the result was a clip just over seven minutes long, and the project is now being called a series. The clip was made with Wonder Studios and director Dev Walker. The timing of future episodes, the costs, and the mix of models used have not been disclosed, and the community criticized the drawn-out exposition, unnatural dialogue, and flat characters.
• Figure AI. Company head Brett Adcock said the design of the Figure 04 humanoid has been finalized. He called it "the first iPhone of humanoids." Its main feature is a hand with near-human dexterity, which has more actuators than the rest of the body combined. Production of components has begun, and the robot's exterior is being kept secret. The robot is being built for the Helix 3.0 model, which Adcock calls "baby AGI."

Data as the new currency of robotics

The most interesting part of the Figure story is the scale of its data collection. The Index app uploads 81 minutes of action video every second. About $8.1 million went into expanding datasets in a single week, and Adcock promises to invest more in data over the course of the year than most robotics startups have raised in their entire history. The bet is clear: a humanoid's competitive advantage will be not the hardware but the volume and quality of the demonstrations that Helix 3.0 learns from.

All of these stories share one thread. Money in AI is moving from standalone models to platforms: subscription-based agents, sovereign models for specific countries, robots built around proprietary datasets. Nvidia wants to charge for software on top of every chip, Sakana is building a national model, and Anthropic is earning billions from a coding agent.

What to watch in the coming months

First, how Anthropic develops Claude Code after the departure of the project's public faces: whether the release pace holds and who becomes the product's new voice. Second, the details of Nvidia's license: if it supplements AI Enterprise rather than replacing it, enterprise customers will get one more bill for every accelerator. Third, spring 2027, when Sakana will show the first version of Sakana Sawara and it will become clear whether the Japanese startup has the resources to train from scratch. Fourth, real demonstrations of Figure 04 and Helix 3.0: for now they are judged only on Adcock's word.

For teams building products on Claude Code, the practical takeaway is simple: follow the official channels, not developers' personal accounts. For those planning enterprise AI factories, it is time to budget not only for hardware but also for a software subscription for agents.