Inherent’s small AI agent just beat bigger models at paper replication. The interesting part is how.

London startup Inherent says its new agent, Faraday, outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on a task that sounds narrow but matters a lot in science: reproducing the findings of published papers without being handed the answer.

That’s a useful benchmark. It says something real.

The company, founded by Google DeepMind alumni and backed by a $50 million seed round, isn’t trying to sell another chatbot that drafts emails and writes tidy summaries. It wants to build what it calls an AI scientist agent, a system that can choose experiments, run them, and learn from the results. The replication task is a test case for that bigger goal.

Why this benchmark matters

Paper replication is one of the few AI evaluation tasks that maps cleanly onto real research work.

A model has to read a paper, infer the method, decide what matters, and figure out whether the reported result can be reproduced. That takes more than pattern matching. It needs judgment. If a system can do that reliably, it’s closer to a research assistant than a text generator.

PhD students do this all the time. Reproducing known results teaches them how to deal with experimental detail, data quirks, and the gap between a clean result on paper and something that holds up in practice. Inherent’s test isn’t random. It’s a decent proxy for whether an agent can do useful scientific work.

There’s still a catch. Replication is only one slice of research. It’s a bounded task with an answer somewhere in the world. Discovery is harder. A model can do well here and still be clumsy at choosing new hypotheses, handling ambiguous evidence, or spotting when a negative result matters.

That gap matters, and a lot of AI companies blur it.

Why the small model matters

The headline detail is not just that Faraday won. It’s that it won using Qwen 3.6 with 27 billion parameters, a relatively small model compared with the frontier systems from Anthropic and OpenAI in the comparison set.

That matters for two reasons.

First, scale still matters, but it doesn’t decide everything. A well-trained 27B model can beat a larger one on a specific task if the surrounding workflow is better. That might mean better prompting, better reward signals, better tool use, or a tighter training loop.

Second, size affects cost. A 27B system is far easier to run, iterate on, and fine-tune than a frontier-scale model. For a startup trying to build a scientific agent that inspects papers, plans experiments, and calls external tools, that matters more than a leaderboard screenshot. Smaller systems are cheaper to serve and easier to operationalize.

The obvious limitation is that one benchmark win doesn’t prove broad superiority. It shows Inherent found a way to make a smaller model work better for one kind of scientific task. That’s promising. It’s not a general rule.

Reinforcement learning seems to be the point

Inherent says the key was training Faraday with reinforcement learning rather than relying mostly on imitation of scientific workflows.

That makes sense.

If you want a model to develop something like taste, you don’t spell out every step in advance. You reward outcomes that look good and let the policy learn which actions produce them. In science, that could mean rewarding an agent for choosing experiments that are informative, efficient, and useful for separating competing explanations.

That’s a different setup from teaching a model to mimic how scientists talk and work. The imitation route often produces systems that sound scientific and don’t do the job well. Reinforcement learning can push a model toward better decisions, but it’s harder to stabilize and harder to evaluate. Reward design gets messy fast. If the reward is too narrow, the agent learns to game it. If it’s too broad, training becomes noisy and expensive.

Inherent looks to be betting that reward-based training will generalize better than teaching the model the etiquette of research. That’s a reasonable bet. It’s also the part most likely to fail if the reward function doesn’t line up with actual scientific value.

The tool choice is telling

One detail stands out: Inherent didn’t build its own coding agent. It had Faraday use OpenAI’s GPT-5.5 Codex instead.

That’s practical, and it says something about how the company is thinking.

A lot of AI startups waste time trying to own every layer of the stack. In science, that can be a distraction. If the goal is to build an agent that can design experiments and reason about results, it makes sense to plug in strong existing code generation or execution tools rather than rebuild them from scratch.

The choice suggests Inherent is acting like a systems company, not a model vanity project. Use the best components where they already exist. Spend the custom training effort on the part that’s actually new: scientific reasoning and experimental judgment.

It also hints at a limit. The company isn’t pretending its agent is a fully autonomous researcher. It’s assembling a workflow. That’s probably the right call. Current agents are still brittle when asked to run end to end without guardrails. They do better when the hardest parts are tightly scoped and the rest is instrumented.

London matters here too

Inherent’s office setup says a bit about the company’s bet. Its dozen employees work in person in King’s Cross, in the same part of London that DeepMind helped turn into a serious AI cluster.

That density matters. Research-heavy startups benefit from being close to people who know how to train models, run experiments, and debug the awkward edge cases that never show up in pitch decks. London doesn’t get enough credit for that. It has built a real concentration of AI talent, even if attention still tilts toward San Francisco.

Hughes has also spoken publicly about ending garden leave in the U.K., which can stop departing employees from joining a competitor for months. That’s not a side issue. It affects how fast talent moves, how quickly startups can hire, and how much friction sits between a researcher and their next job. In AI, where people matter as much as compute, that friction has real cost.

For technical teams deciding where to build, the lesson is straightforward. Research culture, hiring speed, and legal constraints shape execution more than most founders like to admit.

What technical teams should take from this

For engineering leaders and AI builders, Inherent’s result points to a few practical truths.

First, benchmark design matters. If a startup can make a smaller model win on a task that resembles real scientific work, that’s more useful than another leaderboard score on generic reasoning tests. A narrow benchmark can still be meaningful if it captures the workflow you care about.

Second, model choice should follow the job. You don’t always need the biggest model. If the agent works in a well-defined loop with tools, rewards, and checks, a smaller system may be easier to control and much cheaper to run.

Third, “agent” is still a messy word. The useful version is usually a trained system wrapped in a workflow, not a fully autonomous brain. The startup that understands that is already ahead of the ones writing AGI slogans.

Fourth, research agents need more than text generation. They need experiment planning, tool use, failure detection, and a way to tell whether they’re making progress. That’s where the real engineering work sits, and that’s where most companies will struggle.

The big caveat is obvious. Paper replication is impressive, but it’s still a contained problem. Real science is full of bad data, incomplete methods, shifting assumptions, and ugly compromises. A model that can reproduce a published result may still be a long way from one that can propose a genuinely new one.

Still, this is the kind of early signal worth watching. Inherent isn’t promising magic. It’s showing that a smaller model, trained the right way, can do something specific and hard better than larger competitors. That’s a lot more credible than most of the agent noise floating around right now.

If the company can move from replication toward actual hypothesis generation, the benchmark game gets much more interesting.