Faraday AI Agent Beats GPT-5.5 at Replicating Scientific Papers

A small model with a big claim

Inherent is a London AI lab started by former Google DeepMind researchers. It says its new agent, Faraday, did better than much larger systems from OpenAI and Anthropic at one specific job: reproducing the findings of published scientific papers without being given the answers in advance.2 The comparison was against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. Faraday itself runs on Qwen 3.6, a much smaller model with 27 billion parameters.2 The release came only weeks after the company came out of stealth with a $50 million seed round.2

The early coverage mostly agrees on the basic facts. It differs on what they mean. Aggregated reports led with the competitive angle, a small lab beating the giants, and focused on the size difference between Faraday and the frontier systems.13 Those same summaries added a caveat: the result needs independent checking before anyone treats it as settled.13 Inherent's own framing is quieter. Co-founder and chief scientist Edward Hughes said that beating frontier agents was not the main point. What interested the team, he said, was the method they used to build Faraday.2

Our view is that Hughes's emphasis is the right one to take seriously. For companies weighing autonomous agents, the scoreboard matters less than how Faraday was built.

How Faraday was built

Inherent set a higher bar than accuracy. It wanted Faraday to show what the company calls "research taste": a sense of which experiments are worth running and how to design them well.2 To get there, the lab relies mainly on reinforcement learning, which rewards good outcomes instead of spelling out rules. Inherent is betting this approach will carry over to many scientific fields better than training agents on how science is done.2

The most useful detail for enterprise buyers is what Inherent chose not to build. Faraday does not have its own coding tool. It calls OpenAI's GPT-5.5 Codex, much as a human scientist uses existing software instead of writing everything from scratch.2 That makes the "beats GPT-5.5" headline more complicated than it sounds. The winning system is partly built on an OpenAI product. In practice, Faraday is an orchestration layer: a small, specially trained model that decides what to do and hands specialised work to outside tools, including tools from the labs it is being compared with.

This is probably the most important lesson in the story. One analysis argued that if the result holds up, smart agent design and tool coordination may matter as much as raw model size in specialised areas like research automation.1 That idea is already familiar to anyone building agent systems inside large companies. Faraday is a clear public example of it working against named frontier models.

A pattern, not a one-off

Faraday is not the first case this year of a structured agent system beating a bigger general-purpose model on a scientific task. In May, Google DeepMind's AI Co-Mathematician scored 47.9% on the FrontierMath Tier 4 benchmark, solving 23 of 48 problems. That beat the previous record of 39.6%, set by GPT-5.5 Pro.6 DeepMind did not get there with a new foundation model. It used a multi-agent setup in which Gemini 3.1 Pro coordinates specialised agents for literature search, code generation and reasoning, with reviewer agents checking the proofs.6 Separate coverage described that system as a workbench built around how mathematicians actually work, including tracking failed hypotheses.10

The same pattern shows up in academic research. Earlier this month, a team showed that research papers can be turned into AI agents that answer questions about their own methods. Across 300 questions, the converted agents scored 91.2%, compared with 80.3% for Claude working directly from each paper's code. Claude was also slower and more expensive per question.8 The approach has limits: 26 of the 100 biology papers the team tried could not be converted, and it works best on papers that come with usable code.8

Taken together, these results back a clear thesis. In scientific and technical work, the system built around a model is increasingly what sets performance apart. For enterprises, that changes the buying question from "which frontier model?" to "which agent design, on which model, at what cost?"

The reliability problem the headlines skip

There are good reasons not to over-read a single benchmark win. Earlier this year, the ResearchGym benchmark tested a GPT-5-based agent on real AI research tasks and found a sharp gap between what it could do and how reliably it did it. The agent beat the provided baselines in only 1 of 15 runs and completed about 26.5% of sub-tasks on average.9 The researchers listed recurring failures over long tasks: impatience, poor time and resource management, overconfidence in weak hypotheses, and hard limits from context length. Performance usually stopped improving after about nine hours.9 In one run, though, the agent did beat a strong published solution. That shows frontier agents can occasionally reach state-of-the-art results, just not consistently.9

That is the right lens for Faraday. Reproducing a paper is a narrower task than open-ended discovery, and Hughes says so. He describes replication as a stepping stone and notes that many PhD students start their training the same way.2 Inherent's stated aim is much bigger: AI that discovers new scientific knowledge instead of checking old results.2 Companies thinking about deploying research agents should treat Faraday's win as evidence of good design on a well-defined task, not as evidence that autonomous research is solved.

A second caution is timing. GPT-5.5 is no longer OpenAI's top model. OpenAI released GPT-5.6 Sol on July 9, and one public leaderboard puts it well ahead of GPT-5.5.1218 The same tracker shows GPT-5.6 Sol scoring 83% on FrontierMath Tier 4, far above the numbers that made headlines in the spring.14 OpenAI is also aiming directly at enterprise agent workloads. It has an ultrafast GPT-5.6 Sol variant that reportedly runs at 14 times normal speed and is available through an invite-only waitlist.1 Beating GPT-5.5 in August means beating the previous generation, and that edge could shrink quickly.

The enterprise math

The cost argument for a 27-billion-parameter agent core is easy to make, even though Inherent has not published deployment pricing. Model size roughly tracks training cost, and smaller models are generally cheaper to run.2 GPT-5.5 is listed at $5 per million input tokens and $30 per million output tokens.12 Research agents read a lot of material and generate a lot of text over long tasks, so per-token pricing adds up fast. A small, specially trained model that calls an expensive frontier tool only when it needs to is an appealing pattern for organisations trying to keep agent costs down.

Inherent also has a view on how agents should behave. Hughes said the company wants to avoid agents that just tell users what they want to hear. His model is a curious colleague who goes off, runs experiments and comes back asking for an opinion on the results.2 For companies worried about agents that confidently produce plausible but wrong output, that framing goes straight at a real deployment risk. Whether Faraday actually behaves that way consistently is something only outside testing can show.

The DeepMind diaspora context

Inherent's announcement is part of a wider wave of departures from DeepMind. Coverage in late September described venture investors chasing DeepMind alumni as Google moves the centre of its AI work from London to California. That shift was highlighted by Demis Hassabis stepping down in August from running Google's AI and research operations.19 Other ex-DeepMind ventures are raising far more than Inherent. David Silver's Ineffable Intelligence reportedly raised about $1 billion, and Emulate is reportedly in talks to raise up to $700 million at a $3.7 billion valuation.1915

Compared with those, Inherent is small: about a dozen people working in person in King's Cross, with plans to grow to 20 to 25 by the end of the year.2 Reporting suggests its hiring could appeal to DeepMind staff unsettled by the leadership changes.2 That modest size is part of the story. While better-funded rivals have yet to show anything concrete, Inherent is shipping results.2

The verdict

Faraday's win is real but narrow. It is self-reported, limited to research replication, and measured against a model OpenAI has already replaced. What should last is the design approach: a small model trained with reinforcement learning, coordinating outside tools, and judged on research judgment as well as accuracy. For enterprises, the takeaway is not that a startup beat OpenAI. It is that well-designed agents running on modest models can now compete with frontier systems on specialised, high-value work. That argues for judging agent systems on their design, reliability and cost, not on the size of the model underneath.