DeepMind Alumni Startup Inherent Says Its Tiny Agent Outperformed the Frontier Models at Replicating Research

Faraday runs on a 27B-parameter model and beat Opus 4.8 and GPT-5.5 at reproducing published science — with a dollop of "research taste"

Published: 2026-08-23 Category: Quick Take Sources: TechCrunch — Inherent, founded by DeepMind alumni, says its AI 'teammate' just outperformed Anthropic and OpenAI at replicating research

What Happened

London AI lab Inherent, founded by Google DeepMind alumni, says its newly released AI agent Faraday outperformed much larger, better-known models at a specific task: independently reproducing the findings of published scientific papers without being told the answer in advance.

The startup emerged from stealth just weeks ago with a $50 million seed round. What's notable is the efficiency gap. Measured against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 — both frontier-scale systems — Faraday runs on a comparatively tiny model called Qwen 3.6 with just 27 billion parameters.

Why It Matters

Paper replication is a standard training exercise for human scientists — "Many PhD students actually start by doing this," cofounder and chief scientist Edward Hughes told TechCrunch. Inherent's loftier goal is AI that discovers new scientific knowledge, not just verifies old results. But the replication benchmark is a meaningful, verifiable first public demonstration.

Inherent's bar for success was also higher than raw accuracy. Beyond reproducing results, it wanted Faraday to demonstrate "research taste" — an instinct for what experiments are worth running and how to design them well. That's an intangible, and it's where the company's bet on reinforcement learning comes in: rewarding agents for good outcomes rather than spelling out rules, in the hope it generalizes to its longer-term goal of agents that contribute across many scientific fields.

The Deeper Signal

The DeepMind-alumni pedigree matters here. Inherent is far from the flashiest AI startup, but it has chosen a slow, deliberate road — emerging from stealth only to show something concrete. Notably, it also refused to build its own coding tool, having Faraday use OpenAI's GPT-5.5 Codex instead, much the way human scientists lean on existing software rather than reinventing it.

For the broader market, the headline is efficiency: a 27B model beating frontier systems at a hard scientific task upends the assumption that scale is destiny. If reinforcement learning and careful agent design can close the gap on a small model, the economics of AI science agents shift — and the "north star" of a general AI scientist starts looking less like a moonshot and more like a buildable product.

Based on reporting by TechCrunch.