London-based AI startup Inherent announced in August 2026 that its AI agent, Faraday, has outperformed larger models from Anthropic and OpenAI at independently reproducing the findings of published scientific papers — without being told the results in advance.
The result is notable partly because of how Faraday was built. Rather than relying on a frontier-scale model, Inherent used Qwen 3.6, a comparatively small model with 27 billion parameters, to beat Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 at the task. Parameters serve as a rough proxy for model size and training costs.
Inherent was founded by four alumni of Google DeepMind, including cofounder and chief scientist Edward Hughes. The company operates out of King’s Cross in London and employs around a dozen people, with plans to grow to 20 to 25 staff by year’s end. It emerged from stealth just weeks ago with a $50 million seed round.
Hughes said the benchmark result was less important than the method behind it. Inherent trained Faraday using reinforcement learning — a reward-based approach that encourages good outcomes rather than prescribing fixed rules — betting it will generalize across many scientific fields. The company also set a higher bar than accuracy alone, requiring Faraday to demonstrate what it calls “research taste”: an instinct for which experiments are worth running and how to design them well.
Paper replication is a standard exercise for human scientists, Hughes noted. “Many PhD students actually start by doing this.” Inherent’s longer-term goal is building AI capable of discovering new scientific knowledge, not just verifying existing results.
The company has also chosen to avoid reinventing tools that already exist. Rather than building its own coding assistant, Faraday uses OpenAI’s GPT-5.5 Codex — a deliberate design choice Hughes compared to how human scientists rely on existing software.
Faraday’s release could position Inherent as a more visible player in AI research tooling, particularly as it looks to attract talent from larger organizations, including potentially from DeepMind itself.
Source: TechCrunch