DeepMind出身者が設立したInherentのAI「チームメート」、研究再現でAnthropicとOpenAIを上回ると発表Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

British artificial intelligence lab Inherent has launched Faraday, an AI agent built to reproduce the experiments and results described in scientific papers, and the company says it outperforms models from Anthropic and OpenAI on that specific task. The claim matters because reliable, automated reproduction of research could address one of science's most persistent weaknesses and, potentially, help speed up the pace of discovery.

Inherent was founded by alumni of Google DeepMind, the London-based lab behind systems such as AlphaGo and AlphaFold. That lineage places the British startup within a growing cohort of European AI outfits staffed by researchers who trained on large-scale machine learning and reinforcement learning. The company is positioning Faraday less as a chatbot and more as a research "teammate," an agent intended to carry out multi-step technical work rather than simply answer questions.

The core capability Inherent is highlighting is paper replication. In practice, that means reading a scientific paper, understanding its methodology, writing and running the code needed to reproduce its experiments, and checking whether the reported findings hold. This is a demanding test for an AI system because it requires long chains of reasoning, tool use, coding, and the ability to recover from errors, the kinds of tasks that define the current wave of so-called agentic AI.

Reproducibility is a meaningful target. Across fields from psychology to biomedicine, researchers have documented a replication crisis, in which a significant share of published results cannot be independently reproduced. An agent that can reliably attempt replication could help journals, funders, and labs flag fragile findings earlier, and could free human researchers from tedious verification work. Inherent frames Faraday as a possible stepping stone toward broader scientific innovation, though such downstream benefits remain to be demonstrated at scale.

The comparison with Anthropic and OpenAI reflects how the field is increasingly measured. Both companies have released frontier models, Anthropic's Claude family and OpenAI's GPT line, that are frequently evaluated on coding and reasoning benchmarks. OpenAI has itself published work on whether AI can replicate research, including a benchmark called PaperBench that asks models to reproduce machine learning papers from scratch. Inherent's decision to benchmark against these labs suggests it is targeting a similar evaluation space, where the ability to execute a full experimental pipeline, not just generate plausible text, is what counts.

As with most vendor-reported benchmarks, the results should be read with some caution. Performance figures released by a lab about its own system have not necessarily been independently verified, and the details of which papers were tested, how success was scored, and which competitor model versions were used all shape the outcome. Without third-party replication of Inherent's own claims, a fitting irony given the product's purpose, it is difficult to know how the reported advantage generalizes beyond the chosen benchmark.

Faraday also arrives amid a wider industry push to apply AI to scientific work. Google has demonstrated an AI co-scientist system built on its Gemini models to help generate research hypotheses, and DeepMind's AlphaFold has become a widely used tool for protein structure prediction. Startups such as FutureHouse have pursued AI agents aimed at automating parts of the research process. This context suggests that Inherent is entering a competitive and fast-moving area rather than opening an entirely new one, and that credibility in the space is likely to depend on outside validation as much as internal metrics.

For now, the significance of Faraday appears to rest on its narrow but valuable focus: turning the written record of science into reproducible, executable results. If the tool performs as described outside controlled benchmarks, it could become a practical aid for researchers checking prior work or building on it. If it does not, it will still illustrate how the AI industry is shifting its ambitions from conversation toward autonomous, verifiable technical labor. Inherent has not detailed pricing or broad availability in the material released so far, and the practical reach of the system will likely depend on how accessible it becomes to working scientists and on whether independent groups can confirm the performance the company reports.