Inherent, a London AI lab started by former Google DeepMind researchers, says its new agent Faraday has beaten larger models from Anthropic and OpenAI at replicating scientific papers. Faraday runs on a 27-billion-parameter version of Qwen 3.6. In tests it independently reproduced findings from published papers without being given the answers first. It outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on the task. The company also measured “research taste,” meaning the ability to pick useful experiments and design them well. Training relied on reinforcement learning that rewards good outcomes. Faraday still uses OpenAI’s GPT-5.5 Codex for coding work. Inherent raised a $50 million seed round weeks earlier and aims to build agents that can one day discover new science rather than just verify old results.
Inherent, a London AI lab started by former Google DeepMind researchers, says its new agent Faraday has beaten larger models from Anthropic and OpenAI at replicating scientific papers. Faraday runs on a 27-billion-parameter version of Qwen 3.6. In tests it independently reproduced findings from published papers without being given the answers first. It outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on the task. The company also measured “research taste,” meaning the ability to pick useful experiments and design them well. Training relied on reinforcement learning that rewards good outcomes. Faraday still uses OpenAI’s GPT-5.5 Codex for coding work. Inherent raised a $50 million seed round weeks earlier and aims to build agents that can one day discover new science rather than just verify old results.