Picture this: a scrappy startup in London, run by ex-DeepMind researchers, just walked into the heavyweight title fight and knocked out both champions with one punch. Inherent's AI agent, Faraday, outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at replicating published scientific research — and it did it with roughly 27 billion parameters. That's like a moped beating Ferraris on the Autobahn.
Here's the kicker: Faraday wasn't given the answers. It had to reproduce results from scratch, testing real reasoning and experimental design. This isn't a trivia contest. It's a benchmark for genuine scientific thinking. If the claim holds, it's a direct punch to the scaling orthodoxy that OpenAI and Anthropic have bet their valuations on — the idea that bigger is always better.
But here's the twist: Faraday uses rewards, not rules, and focuses on 'research taste.' That's a different philosophy entirely. It's less about memorizing patterns and more about knowing what's worth testing. It's a reminder that intelligence isn't just about compute — it's about judgment. And that's a threat to the narrative that frontier AI requires frontier-scale infrastructure.
The losers are obvious: OpenAI and Anthropic have spent years convincing the world that massive data centers and billions in capex are the only path forward. If a small model can do the work of giants, those moats start looking like sunk costs. And for Google DeepMind, the talent drain just got more painful — its own alumni are now beating its rivals.
The winners? Inherent, obviously. But also the scientific community, which could soon have an AI that actually does science, not just summarizes it. And any enterprise tired of paying for bloated models when a leaner one might suffice.
So what does this mean for the average reader? It means the AI race might not be about who has the deepest pockets, but who has the smartest ideas. The capex race could be a mistake — and trends suggest we're entering an era of efficiency-first AI. Could Faraday be the start of a new wave of 'small but mighty' models? Or will the giants absorb this lesson and adapt? Either way, the status quo just got a serious wake-up call.