27B-Parameter Faraday Beats OpenAI and Anthropic in Lab Test
A 27-billion-parameter AI from $50 million startup Inherent beat Claude Opus 4.8 and GPT-5.5 at reproducing published results on its own in tests.
Summary
A 27-billion-parameter AI just beat much larger frontier models at reproducing published science. On August 22, 2026, London startup Inherent said Faraday, running on Qwen 3.6, outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 without seeing answers first.
Fresh from stealth with a $50 million seed round, the DeepMind alumni lab uses reinforcement learning to teach “research taste,” while Faraday taps GPT-5.5 Codex for coding. Edward Hughes founded Inherent with Louis Kirsch, Kaloyan Aleksiev and Tantum Collins. Its 12-person King’s Cross team plans to reach about 20 to 25 employees by the end of 2026. The bigger goal is AI that discovers new scientific knowledge.
Positives
- Faraday beat much larger frontier models using Qwen 3.6 with 27 billion parameters.
- A $50 million seed round gives Inherent fuel to chase AI-driven scientific discovery.
- The 12-person team plans to grow to about 20 to 25 employees in 2026.
Risks & concerns
- The victory covers replication, not Faraday discovering new scientific knowledge.
- Faraday still relies on OpenAI’s GPT-5.5 Codex for coding.
- U.K. garden leave restrictions can delay researchers joining or launching rival AI labs.