Inherent, a London AI lab founded by Google DeepMind alumni, has released an AI agent called Faraday that the company says outperformed larger systems from Anthropic and OpenAI on a scientific-paper replication task, according to TechCrunch. The task was specific: independently reproduce the findings of published scientific papers without being told the answer in advance. TechCrunch reports that Inherent frames this as a step toward a larger goal — building AI systems that can contribute to scientific discovery, not merely verify existing results. The core claim remains company-reported. Inherent told TechCrunch that Faraday beat Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on the replication benchmark. The cluster does not include an independent evaluation of that result, so the performance claim should be read as Inherent’s account rather than a settled external ranking. The size comparison is the more concrete technical hook. TechCrunch reports that Faraday runs on Qwen 3.6, a 27 billion-parameter model, while it was measured against much larger frontier-scale systems. Parameters are a rough proxy for model size and often for training cost, though they do not alone determine capability. Inherent’s stated bar was not only whether the agent could reproduce a paper’s results. The company wanted Faraday to show what cofounder and chief scientist Edward Hughes described to TechCrunch as “research taste”: judgment about which experiments are worth running and how to design them. Hughes compared the replication task to a common early exercise for human researchers, saying many PhD students begin by reproducing prior work. TechCrunch reports that reinforcement learning is central to Inherent’s approach. Rather than giving the system fixed rules for how to conduct science, the company rewards outcomes it considers useful. Inherent’s bet, as described by TechCrunch, is that this reward-driven method will generalize better as the company works toward agents that can operate across scientific domains. The company is also drawing a boundary around what it does not want to build. According to TechCrunch, Inherent did not create its own coding tool for Faraday; it had the agent use OpenAI’s GPT-5.5 Codex instead. The company likens that to human scientists using existing software tools rather than rebuilding every component themselves. The timing matters because TechCrunch says Inherent emerged from stealth only weeks ago with a $50 million seed round. Among startups founded by DeepMind alumni, TechCrunch says Inherent has received relatively little attention and that better-funded rivals have yet to show the world anything concrete; Faraday is how the London-based team is starting to share what it has been building, with a focus on scientific workflow skills rather than general chatbot use. The unanswered question is validation. A single company-reported benchmark, even when described in detail by a reputable outlet, does not establish that Faraday is broadly better than frontier models at scientific work. What it does establish is Inherent’s thesis: smaller models, agent scaffolding, reinforcement learning, and tool use may be enough to compete on narrow research tasks if the system is trained around the right objective. Who benefits: Inherent benefits from showing a concrete system soon after its reported $50 million seed round. Scientific teams that need help reproducing papers could also benefit if Faraday’s reported capabilities hold up outside the company’s own evaluation. Who's exposed: Large-model providers are exposed only in a limited sense: the claim concerns one paper-replication task, not general model superiority. Labs building broad “AI scientist” systems may face pressure to show similar task-level evidence rather than high-level ambition.