Small 27B AI Model Beats Claude Opus 4.8 and GPT-5.5 at Replicating Science
News Summary
A London-based AI lab called Inherent, founded by a group of former Google DeepMind researchers, announced on August 22, 2026, Eastern Time, that its AI research agent, Faraday, had outperformed frontier-scale systems from Anthropic and OpenAI on a difficult new benchmark: independently reproducing the results of published scientific papers without being told the answer in advance. The claim stands out not just for the win itself, but for the size gap behind it — Faraday runs on a comparatively small 27-billion-parameter model, while the systems it beat, Claude Opus 4.8 and GPT-5.5, are both far larger, general-purpose frontier models.
Who Is Inherent
Inherent emerged from stealth in May 2026 with a $50 million seed round. The company was founded by four cofounders, three of whom previously worked at Google DeepMind: Edward Hughes, who serves as chief scientist, along with Louis Kirsch and Kaloyan Aleksiev, plus Tantum Collins. The team is based in London's King's Cross neighborhood and currently numbers around a dozen employees, with plans to grow to 20 to 25 staff by the end of the year. Rather than building its own coding tools from scratch, Inherent relies on OpenAI's GPT-5.5 Codex to handle raw coding tasks, choosing instead to focus its research effort on a narrower and, in the founders' view, more important problem.
The Core Idea: Teaching "Research Taste"
Hughes has described the company's central bet as teaching AI systems "research taste" — the instinct for which experiments are worth running, how to design them well, and when a result is believable. In Inherent's framing, most existing AI coding agents are good at executing well-specified instructions but weak at the judgment calls that precede and follow execution: forming a hypothesis, deciding what to test first, and evaluating whether an outcome makes sense. According to the company, Faraday is built to fill that gap by acting as a "director" that delegates the mechanical coding work to a separate coding model while it concentrates on the higher-level scientific reasoning.
How Faraday Was Tested: The Replica Benchmark
To evaluate this idea, Inherent built a new benchmark called Replica, made up of 310 tasks drawn from 100 machine learning and AI-for-science papers covering fields such as natural language processing, materials science, and weather forecasting. Each task takes a published paper, strips out one of its results figures, and asks the agent to recreate that result by actually running the underlying experiment, without access to the original plot and under limited time and compute budgets. Because research papers typically describe only what ultimately worked, and omit the failed attempts, missing hyperparameters, and preprocessing steps that led there, the benchmark is designed to test whether an agent can recover what the Inherent team calls the "99% perspiration" that never makes it into the final paper.
Faraday was trained on 242 replication tasks and then evaluated on 68 held-out tasks it had not seen during training. On that held-out set, Faraday scored 0.791, ahead of Claude Opus 4.8 at 0.748 and GPT-5.5 at 0.729, and won outright on 60 percent of the individual tasks. Inherent reports that Faraday produced more faithful replications than both larger models across every category in the Replica suite, with the widest margins showing up in meta-learning, structural biology, and materials science tasks.
The Training Approach
Faraday was trained using long-horizon reinforcement learning, with a general-purpose coding agent acting as a tool it directs rather than a component it replaces. To keep that training stable, Inherent's researchers used per-task scoring rubrics instead of relying solely on an AI judge to grade outcomes, since rubric-based scoring produced a more consistent and less noisy reward signal. The team also aggregated multiple samples during training and applied turn-level credit assignment, which helps the model learn from long sequences of actions rather than only from a single final outcome. Notably, Inherent says Faraday was trained while directing an earlier, smaller coding model, GPT-5.4-mini, but was still able to generalize its skills to direct the more capable GPT-5.5 Codex at test time without additional retraining — evidence, the company argues, that the "director" skill it is teaching is somewhat independent of which underlying coding tool is available.
Context and What It Signals
Inherent has positioned the Replica benchmark and Faraday's performance on it as an early proof point for its broader ambition: building AI systems that can meaningfully accelerate scientific discovery, rather than simply automate coding tasks. The company is not alone in pursuing this direction — automating parts of the AI research process itself has become an active area of interest across the field, as labs explore how much of experiment design, execution, and evaluation can be handed to AI systems. Inherent's specific contribution is a benchmark and training recipe aimed squarely at the judgment-heavy parts of research that are hardest to specify with simple instructions, and an early result suggesting that a smaller, more narrowly trained model can match or exceed much larger general-purpose systems on that particular skill.