OpenAI Just Ate Perplexity's Lunch and Called It a Dev Tool

The announcement came in pieces, which is how you know it's actually significant — if it were just a feature drop they'd have put it in a blog post with a gradient header.

Web search is now a native tool in the Agents API. You call it through the OpenAI API, it returns results, you pay 4o-mini prices. Which means the entire pitch for paying Perplexity $5 per thousand searches just collapsed into a line item that rounds to zero.

Perplexity is a great product. The answer interface is genuinely good. But "LLM that searches the web" was always a borrowed moat, and OpenAI just built the moat into the water table.

File search too. And computer use — they're calling it Operator in the consumer product, but the API surface gives you the same thing: point it at a machine, it clicks buttons and fills forms. Any machine. The "any" contains multitudes and I'm not sure anyone has fully thought through what that means at scale, but sure, let's ship it.

The part that actually caught me off guard: Swarm is becoming the Agents SDK.

Swarm was released last year as an "educational framework" — which is the polite way of saying "we're not supporting this, don't use it in prod, here's the repo." It was a sketch. A proof of concept for how you'd orchestrate multiple agents, hand off context between them, run them in parallel.

Now it's a production framework. Same ideas, different SLA.

The timeline on that is fast enough to be disorienting — experimental to production in under six months, for something that coordinates autonomous agents running with tool access on live machines. The decision to ship Swarm as a research preview first was either a very good testing strategy or a very good PR strategy, and honestly it doesn't matter which.

What matters is that the pieces are all here now. Search, file access, computer control, multi-agent orchestration. They fit together into something that can run a workflow end-to-end without a human in the loop.

Whether that's exciting or terrifying is mostly a function of which workflows you're thinking about.