Inherent, a London lab founded by Google DeepMind alumni, says its Faraday agent replicated published scientific findings better than larger systems from Anthropic and OpenAI. The startup told TechCrunch that Faraday uses Qwen 3.6, a comparatively small 27-billion-parameter model, while its benchmark comparisons used Claude Opus 4.8 and GPT-5.5.
The task was more demanding than producing a plausible summary. Faraday had to reproduce results from scientific papers without being given the answer, and Inherent also evaluated “research taste”: choosing worthwhile experiments and designing them sensibly. The company emerged from stealth with a $50 million seed round and plans to grow from about a dozen employees to 20–25 by year-end. Inherent says Faraday uses OpenAI's GPT-5.5 Codex for coding, a reminder that an agent's advantage can come from orchestrating existing tools rather than replacing every component. The result is promising, but the company has not yet shown that the same method transfers cleanly beyond its chosen task.
Unique insight
Faraday's significance is not that a small model beat larger rivals once; it is that reinforcement learning may make specialised agents more economical than ever-larger general models.
Analysis
If the result survives independent scrutiny, it changes the economics of scientific AI. Labs may not need frontier-scale models for every research workflow; a smaller system with the right tools, reward design and autonomy could deliver more useful work per dollar. Researchers are affected first, because the agent is aimed at experiment selection and replication rather than conversational assistance. The claim remains a company-reported result, so the key question is reproducibility across fields and datasets, not the headline leaderboard. Watch for a technical report, public evaluations and evidence that Faraday can generate genuinely new findings rather than reliably re-run known experiments.
Bottom line — If I only remember one thing tomorrow, what should it be?
The next research-agent advantage may come from specialised training and judgment, not from simply scaling the underlying model.
- Story type: Important story
- Impact: High
- Confidence: Medium (multiple reports)
- Audience: Researchers
- What changed: Inherent released Faraday and reported superior paper-replication results using a 27-billion-parameter model.
- What to watch: Independent tests of Faraday on unseen scientific tasks.
- Reading time: 2 minutes
