Inherent, a London AI lab founded by Google DeepMind alumni, said its Faraday agent outperformed Claude Opus 4.8 and GPT-5.5 at independently reproducing the findings of published scientific papers, using a model a fraction of the size of either frontier system. Faraday runs on Qwen 3.6, which carries just 27 billion parameters — a proxy for both model size and training cost that puts it well below the scale of the systems it beat. The result arrives weeks after Inherent left stealth with a $50 million seed round.
The task itself is standard training work for early-career human scientists. Paper replication means taking a published study, ignoring the stated result, and running the experiments to see if the numbers hold. Inherent cofounder and chief scientist Edward Hughes said many PhD students start their careers doing exactly this, which makes it a useful proving ground for an AI agent aimed at a much larger goal — actual scientific discovery.
Hughes said the benchmark result matters less than the method behind it. Inherent's bet is that its training approach, not raw parameter count, is what closed the gap with Anthropic and OpenAI.
“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.”— Edward Hughes, Inherent cofounder and chief scientist
Key facts
- 01Inherent's Faraday agent runs on Qwen 3.6 with 27 billion parameters and beat Claude Opus 4.8 and GPT-5.5 at replicating published research findings.
- 02The London startup emerged from stealth weeks ago with a $50 million seed round and now employs a dozen people out of King's Cross.
- 03Inherent plans to grow headcount to roughly 20 to 25 employees by the end of the year, targeting DeepMind talent among others.
- 04Faraday uses OpenAI's GPT-5.5 Codex as its coding tool rather than building an in-house equivalent.
- 05Cofounder Edward Hughes is publicly pushing to end UK 'garden leave' clauses that delay AI researchers moving between labs.
The company leaned on reinforcement learning rather than training Faraday primarily on the mechanics of how science is conducted. Reinforcement learning rewards models for good outcomes instead of encoding explicit rules, and Inherent is wagering that the approach generalizes better across scientific fields. The target is what the company calls research taste — an instinct for which experiments are worth running and how to design them well.
That framing has shaped what Inherent has chosen not to build. Rather than developing its own coding tool, Faraday calls OpenAI's GPT-5.5 Codex, on the logic that human scientists also rely on existing software rather than reinventing every tool from scratch. It's a deliberate narrowing of scope for a small team.
The company is also trying to avoid building an agent that flatters its user.
Hughes described the design goal as an agent that behaves like a curious teammate — one that comes back with unsolicited experiments and asks what the user thinks of the results, rather than one that produces confident-sounding summaries on demand. The push against sycophancy echoes a broader concern across frontier labs, where models optimized on human feedback have repeatedly drifted toward agreeable rather than accurate responses.
Inherent has a dozen employees, all working in person out of a King's Cross office in the London neighborhood that Google DeepMind's presence helped establish as a global AI hub. Hughes said Inherent plans to grow to roughly 20 to 25 people by year-end, and with Demis Hassabis taking on a new role at DeepMind that has unsettled some staff there, Inherent's hiring push may find willing candidates nearby.
Hughes has also added his voice to calls to end UK garden-leave clauses, which bar departing employees from joining or founding rival companies for months after resigning. He said he personally ran into the constraint before starting Inherent with cofounders Louis Kirsch, Kaloyan Aleksiev and Tantum Collins.
“This is a personal view rather than a company view, but I was affected by the garden leave problem.”— Edward Hughes, Inherent cofounder and chief scientist
The claim to beat frontier labs on a research-replication benchmark comes with caveats. Inherent has not published a peer-reviewed evaluation, the exact benchmark suite and scoring methodology have not been detailed publicly, and Anthropic and OpenAI have not responded to the specific comparison. Faraday also outsources coding to GPT-5.5 Codex, which complicates any clean parameter-to-parameter comparison against the systems it says it outperformed.
For a $50 million seed-stage lab to claim any win over Claude Opus 4.8 and GPT-5.5 is the sort of result that either holds up under external scrutiny or quietly evaporates. If Inherent's reinforcement-learning-on-taste approach genuinely produces frontier-competitive agents at 27 billion parameters, the implication for the AI-model market is significant: the cost curve for research-grade agents drops sharply, and the moat around trillion-parameter frontier systems narrows to specific capabilities rather than general reasoning. The company now has to show the work.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
