Inherent Claims 27B Agent Beat Claude, GPT-5.5 at Research

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 113 reads

London AI lab Inherent, founded by Google DeepMind alumni, says its newly released agent Faraday has outperformed both Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on a scientific paper-replication task - despite running on a base model, Qwen 3.6, with just 27 billion parameters, a fraction of the scale of those frontier systems.

The claim, shared publicly just weeks after the company emerged from stealth with a $50 million seed round, is the first concrete technical benchmark from one of the more quietly watched startups to come out of the DeepMind diaspora. The task Faraday was tested on is specific: given a published scientific paper, reproduce its findings independently, without being told the answer in advance.

The result sits at an interesting intersection of model efficiency and scientific capability - two of the most commercially valuable properties an AI lab can demonstrate right now.

Why Paper Replication Is the Right Test

Scientific paper replication might sound like a narrow or even underwhelming benchmark for a lab whose stated ambition is to discover new scientific knowledge. Cofounder and chief scientist Edward Hughes is aware of that perception, and pushes back on it directly.

“Many PhD students actually start by doing this.”

Hughes told TechCrunch that the point of the benchmark was not primarily about beating competitors. “What was most interesting to us about this was not so much the result of beating those frontier agents - which of course we liked - but was actually the way we went about building this,” he said.

That distinction matters for how the result should be read. Replication is a standard scientific training exercise because it forces a researcher to understand methodology, not just conclusions. For an AI agent, doing it without prior knowledge of the answer tests a kind of independent reasoning that straightforward question-answering benchmarks do not. It is also reproducible and verifiable, which makes it a credible, if limited, proxy for broader scientific reasoning.

Inherent set a higher bar than raw accuracy. Beyond getting results right, it wanted Faraday to demonstrate what Hughes calls “research taste” - an instinct for which experiments are worth running and how to design them well. That is a considerably harder property to measure or teach than factual recall.

Reinforcement Learning as the Core Bet

The mechanism Inherent used to instill that taste is reinforcement learning, a training approach that rewards an AI system for good outcomes rather than encoding explicit rules. Instead of training Faraday primarily on meta-level study of how science is conducted - reading papers about scientific method, for instance - the company leaned on reward-based learning, betting it will generalize more robustly to the long-term goal of agents that can contribute across many scientific fields.

The choice to build on Qwen 3.6 at 27 billion parameters, rather than training a larger proprietary model, is itself a strategic signal. In rough terms, parameter count is a proxy for both model size and training cost. Matching or exceeding much larger systems on a meaningful task at that scale is a strong efficiency argument - the kind that matters for deployment economics, not just academic bragging rights.

Inherent also made a deliberate choice about what not to build. Rather than developing an in-house coding tool, the company had Faraday use OpenAI’s GPT-5.5 Codex for code generation - the same way working scientists reach for existing software rather than writing everything from scratch. It is a pragmatic division of labor, and it keeps the team focused on the higher-level reasoning layer where Inherent believes its edge lies.

What This Means for the Market

The efficiency story has clear commercial implications. A 27-billion-parameter agent that matches frontier performance on scientific reasoning tasks costs far less to run at scale than GPT-5.5 or Claude Opus 4.8. For pharmaceutical companies, materials science labs, or any enterprise running large volumes of research workflows, that cost difference compounds quickly.

The broader vision Hughes describes is of an AI that behaves less like a search engine and more like a genuine research collaborator - one that goes away, runs experiments autonomously, and returns with something unexpected. “I got curious about this, and I went off and I did these experiments. What do you think of these results?” is how he framed the ideal interaction.

Inherent’s dozen employees all work in person from an office in King’s Cross, the London neighborhood whose transformation into a leading AI hub was accelerated by Google DeepMind’s own presence there. The lab is still early, and a single benchmark on a specific task is not a general proof of scientific discovery capability. But in a field where most DeepMind spin-outs have announced ambitions well ahead of results, a concrete, named benchmark with specific model comparisons is a more grounded opening move than most.

Whether Faraday can make the much harder leap from replication to genuine discovery is the question that will define whether Inherent’s $50 million seed round looks prescient or premature.

// read moreClaude Code Agent in Pentagon System: WIRED ClaimsNewsClaude Opus 5.5 Data Analysis: 43 of 44 Claims VerifiedBenchmarksAI Agents Run Experiments but Fail Original Research - PrincetonNews

Kinetyra