Startup Inherent: the small Faraday model beat Claude and GPT at reproducing scientific papers

Inherent, a London-based lab founded by Google DeepMind alumni, has announced that its AI agent Faraday beat much larger models from Anthropic and OpenAI at a specific task — independently reproducing the conclusions of published scientific papers without knowing the answer in advance. Just a few weeks earlier the startup had emerged from stealth with a $50 million seed round.

The comparison was run against Claude Opus 4.8 and GPT-5.5 — frontier-scale systems. Yet Faraday runs on the comparatively small Qwen 3.6 model with 27 billion parameters, while its rivals are considerably larger (the parameter count is an indirect reflection of both a model's size and the cost of training it).

Not the result, but the method

Co-founder and chief scientist Edward Hughes stresses that what mattered more than the win itself was the way it was achieved. Reproducing others' work is a standard exercise for human scientists too: it's often where PhD students begin. But at Inherent they set the bar higher than mere accuracy: they wanted Faraday to have “research taste” — an instinct for which experiments are worth running and how to design them well.

This elusive quality, according to Hughes, is taught through reinforcement learning: the system is rewarded for good outcomes rather than being handed rigid rules. The company is betting precisely on this approach, expecting it to generalize better toward the ultimate goal — a scientist agent capable of working across different scientific fields.

The “colleague” philosophy

Inherent deliberately doesn't build everything itself. Instead of its own tool for writing code, Faraday uses OpenAI's GPT-5.5 Codex — just as human researchers rely on off-the-shelf software. Nor does the company want a sycophantic assistant that echoes the user. Hughes's ideal is a colleague who comes back and says: “I got curious, I ran these experiments — what do you make of the results?”

The entire team of a dozen employees works in person at an office in the King's Cross area, which DeepMind's presence has turned into one of the world's leading AI hubs. By the end of the year they plan to grow the headcount to 20–25 people.

Separately, Hughes — speaking for himself, not the company — came out against the British practice of garden leave, in which a departing employee is barred for months from joining competitors or launching their own venture. American researchers usually face no such restriction, which gives U.S. startups a head start in hiring. Hughes admits he ran into this problem himself, but ultimately got around it and founded Inherent together with three other co-founders.

Source: TechCrunch