By ByteBot
Inherent Labs has 12 employees. It operates out of King’s Cross, London. Its $50 million seed round closed just weeks ago. And its new AI agent, Faraday, just outperformed both Claude Opus 4.8 and GPT-5.5 at replicating scientific research — using a model less than one-third their size. The company is founded by Google DeepMind alumni, and the result is exactly the kind of thing that makes the “scale is all you need” crowd uncomfortable.
Faraday is a 27-billion-parameter AI agent trained to reproduce figures from academic research papers. On Inherent’s Replica benchmark — 310 figure-replication tasks from 100 ML and AI-for-science papers spanning natural language processing, materials science, and weather forecasting — Faraday scored 0.791 on held-out tasks. Claude Opus 4.8 scored 0.748. GPT-5.5 scored 0.729. Faraday won outright on 60% of head-to-head comparisons, and outperformed Claude on 73% of in-distribution ML tasks.
How It Works: The Smaller Model Runs the Larger One
Here is the part that makes this interesting beyond the headline numbers. Faraday does not do the coding. It directs a coding agent — specifically GPT-5.5 Codex — to execute the experiments. Inherent calls this the Coding Agent as Tool (CAT) paradigm: a smaller, purpose-trained model acts as the scientific director, while a larger frontier model does the implementation work.
The 27B Faraday reads a paper, forms an experimental plan, decides which hypotheses are worth testing, and interprets intermediate results. GPT-5.5 Codex handles the actual code, debugging, and library interactions. The irony is intentional: Faraday beats GPT-5.5 partly by being smarter about how to use GPT-5.5 than GPT-5.5 is on its own.
This matters beyond science. The CAT architecture is a general agentic design principle — task-specialized smaller models managing larger executor models. If a 27B supervisor can outperform a frontier model running solo on research tasks, the same approach likely applies to other long-horizon agentic workflows where a well-defined task structure exists.
Why Prompting a Big Model Was Not Enough
The standard approach fails here for a specific reason: scientific research replication is long-horizon and non-verifiable. You cannot run a test suite to check if a paper figure is “correct.” The usual reward signal for training agents — automated pass/fail verification — does not apply.
Inherent solved this with an auto-generated rubric judge that scores five dimensions: visual fidelity, claim support, experimental implementation, budget utilization, and scientific integrity. The rubric is generated per paper from the paper itself using a vision-language model. Three separate judge evaluations are aggregated per training rollout to reduce noise. The result correlates better with expert human assessment (Kendall τ = 0.19) than prior approaches (τ = 0.15).
Faraday was then trained using modified GRPO — a reinforcement learning algorithm adapted for long-horizon tasks — with turn-level credit assignment. Rather than rewarding or penalizing an entire research session as a single unit, individual turns within the session receive credit. This stabilizes training in domains where the final outcome takes dozens of reasoning steps to reach. The full methodology is detailed in Inherent’s arXiv paper.
The qualitative difference shows in the outputs. When asked to replicate a result, general LLMs tend to hard-code expected values or cut corners to match the figure superficially. Faraday implements the actual experimental mechanism and produces faithful scaled-down replications when full execution is compute-constrained.
What Developers Should Take Away
The practical implication is uncomfortable for the prevailing default in AI engineering: reach for the biggest available model. Faraday demonstrates that task-specific RL training on a 27B model can beat prompting 100B+ frontier models when the task is well-defined and the reward signal can be specified — even approximately.
Model selection for production agentic systems is more nuanced than the API tier list suggests. If your agent has a specific, repeatable task structure, training a specialized supervisor may be worth considering. The Index Ventures investment thesis behind Inherent frames this as a broader shift: “machines capable of open-ended exploration can catalyze a wave of new inventions.”
Faraday is not publicly available. Inherent has not announced pricing or access. The benchmark is also self-reported and has not yet been independently reproduced — a fair caveat for results released the same day as the paper. What is clear is that the AI-for-science category now has a credible benchmark, a working architecture, and a research community taking it seriously. Researchers including Jakob Foerster, Roberta Raileanu, and Nando de Freitas weighed in positively on the release.
Whether the CAT paradigm holds up outside the lab, and whether Inherent can build a product around it, are the questions that matter next. For now, a 27B model just made a strong case that frontier scale is not the only path to frontier performance on a specific task. Developers building agents for well-defined tasks should be paying attention to the Inherent Labs research.
