Small AI Scientist Beats Frontier Models on Research Tasks

A London startup says its AI scientist agent can reproduce published research results more effectively than much larger frontier systems, pointing to a shift in how builders should think about agent performance.

What happened

Inherent Labs told TechCrunch that Faraday, its research agent, outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 on a benchmark that asks systems to independently reproduce findings from scientific papers. The company says Faraday is built around Qwen 3.6, a comparatively small 27-billion-parameter model.

The result is not a claim that a 27B model is broadly smarter than frontier models. Faraday targets a defined workflow: reading a paper, deciding which experiments are worth running, directing code execution, and judging whether the result faithfully reproduces the original finding. Inherent says reinforcement learning helps train this combination of technical execution and what it calls research taste.

Why it matters for builders

Faraday's architecture separates decision-making from execution. The smaller model acts as a scientific supervisor, while OpenAI's GPT-5.5 Codex handles coding and debugging. That is a useful pattern for production systems: use a specialized model to plan and evaluate, then call a stronger general model only where implementation complexity requires it.

The same principle appears in other agent stacks, including the real-device workflows discussed in n8n Lab's Qwen UI-Agent analysis: the model is only one part of the system. Scaffolding, tools, memory, validation, and task-specific training can change the outcome more than parameter count alone.

For teams building long-running research, coding, or operations agents, the practical takeaway is to benchmark the complete workflow rather than compare model leaderboards. Define what a successful run means, measure recovery from failed tool calls, and give the system a verifier that can reject plausible but incorrect work.

Inherent has not released Faraday as a public product, and the reported benchmark results have not yet been independently reproduced. Still, the direction is important. Specialized supervisors paired with general-purpose executors could make capable agents cheaper, easier to control, and better aligned with the structure of the work they perform.