Inherent, a London-based AI lab founded by former DeepMind employees, has introduced Faraday, an AI agent designed to independently reproduce published scientific experiments. Faraday is based on Qwen 3.6, a 27-billion-parameter model trained with reinforcement learning to develop scientific judgment—deciding which experiments to run, how to design them, and when to stop. Faraday delegates actual code generation to GPT-5.5 Codex.
What Faraday is and how it was evaluated
- Model backbone: Qwen 3.6, 27 billion parameters.
- Training approach: reinforcement learning focused on experimental planning and decision-making.
- Division of labor: Faraday handles experiment design and strategic choices; GPT-5.5 Codex performs code generation.
- Reported performance: on Inherent’s internal benchmark, Faraday outperformed purely using Claude Opus 4.8 and GPT-5.5 Codex for reproducing experimental results.
The comparisons reported by Inherent rely on their in-house benchmark; the announcement does not cite independent third-party validation or fully public benchmark datasets and metrics.
Why this matters
Many AI labs have invested heavily in ever-larger models and more compute to push frontier capabilities. Faraday presents a counterexample: a smaller model that acts as an effective manager of a stronger coding agent can, according to Inherent’s results, achieve better outcomes in reproducing experiments than larger or more expensive models alone.
This suggests that some resources may have been allocated suboptimally across the stack—effort has often gone into making the "worker" models more capable, while the coordinating "manager" role received less focus. If Faraday’s advantages generalize beyond Inherent’s tests and are independently replicated, it could prompt a reassessment of where labs invest in scale and architecture.
Limitations and next steps
- Results so far are reported on Inherent’s internal benchmark; independent external evaluation has not been reported.
- The announcement does not fully disclose the benchmark metrics, the specific tasks used, or the difficulty of the reproduced experiments.
- Technical details about Faraday’s decision-making mechanisms and its interaction protocol with GPT-5.5 Codex will be needed for broader community vetting.
Conclusion
Inherent’s Faraday demonstrates that a 27-billion-parameter model trained for scientific judgment and acting as an orchestration layer, while delegating code generation to a stronger model, can outperform other approaches on the company’s benchmark for reproducing published experiments. The result spotlights the potential value of investing in smaller, specialized manager models that supervise powerful workers rather than only scaling monolithic frontier models.
