OpenAI vs. Anthropic: The Battle Lines Are Drawn

Live from Code with Claude: a conference with real stakes

The hosts open from Anthropic’s developer conference, “Code with Claude,” with Kieran, GM of Kora and, as they jokingly put it, the “father of compound engineering.” The vibe is playful — they promise updates on what they saw, what they didn’t hear, and even a food review — but the main takeaway lands fast: this event matters because Anthropic’s position versus OpenAI is changing.

The Colossus deal could fix Anthropic’s biggest weakness

The day’s biggest news, in their view, is Anthropic’s deal with xAI and SpaceX to tap “all of the compute” in SpaceX’s Colossus cluster. They frame that as a huge shift for a company that’s felt compute-constrained during the rise of the “Claude Code movement,” including unpopular restrictions like blocking cloud subscriptions from OpenClaw. The immediate user-facing result: about 2x usage limits and looser windows, which they describe as the kind of boring improvement that feels amazing if you actually use the product every day.

No “Mythos” reveal — but the platform launch may matter more

Stepping back, the hosts say Anthropic didn’t unveil some massive cinematic model drop — “they launched Mythos or whatever” is the running joke. Instead, the biggest actual launch was Claude managed agents, with memory, orchestration, and outcome-based execution, meaning you can assign an agent a goal and have it return only when the work is done. Their point is that these “sneaky” launches often look modest at first and then end up reshaping how people build.

The battle lines are becoming clearer

One of the sharpest observations in the video is that competition is splitting into two fronts. On one side is the “work operating system” layer — if you’re doing knowledge work on a computer, they think you’ll increasingly do it in something like Claude Code or Codex. On the other is the cloud layer for long-running, asynchronous, team-facing agents, and Anthropic’s managed agents product looks like an early attempt to own that second category before OpenAI has a true hosted equivalent.

Why managed agents feel bigger than they sound

Kieran says building this infrastructure yourself still means reinventing the wheel, so a managed service that is scalable and reliable solves a very real pain. He asks why you wouldn’t just run this on a Mac Mini, then answers his own question: a Mac Mini can go down, but bigger companies need reliability, and “getting 90% there is easy” while getting it to work all the time is the hard, valuable part. That’s why he says he’d use Anthropic’s version immediately if it really delivers dependable execution.

Anthropic’s edge: they seem furthest along on orchestration

Pressed on why he thinks Anthropic is best at agent orchestration, Kieran points to ideas already visible in Claude products: dispatch, agent teams, and the pattern where one agent spins up several others. He says that paradigm — one manager agent coordinating many workers — is going to get big this year, and credits people like Daisy at Anthropic for already building the systems behind it. His argument isn’t just that Anthropic has features; it’s that the people thinking furthest ahead about multi-agent coordination are already there.

The food review, and the final verdict

The required conference food check lands on a “teriyaki glazed salmon bowl” that apparently had no detectable teriyaki, though it did score points for feeling healthy. They rank it above Microsoft Build’s sad Saran-wrapped turkey sandwich, but still middle of the pack overall. Then they close by restating the two big takeaways: the compute deal is a major competitive move, and managed agents are “sneaky important” enough that they expect to be talking about them for months or years.