AI scientist agents have just produced their most provocative result of the year. On 22 August 2026, TechCrunch reported that Inherent — a London lab founded by Google DeepMind alumni — says its AI “teammate”, a 27-billion-parameter agent called Faraday, outperformed frontier systems from Anthropic and OpenAI at replicating the findings of published research papers. The claim comes from Inherent’s own paper, released on arXiv on 13 August, and the numbers behind it are specific: Faraday beat Claude Opus 4.8 and GPT-5.5 Codex on 73% of in-distribution tasks and 60% of held-out AI-for-science tasks.
Replication is a pointed choice of battlefield. Science has spent a decade wrestling with a replication crisis, and reproducing a paper’s figures from its methods is exactly the slow, careful work that human researchers rarely get credit for. Inherent’s bet is that an AI scientist which learns to replicate results — without peeking at the answers — is learning the scientific judgement that raw coding ability lacks. Many PhD students, as co-founder Edward Hughes told TechCrunch, actually start by doing this.
This article covers what artificial intelligence teams and business leaders should take from the announcement: what Faraday is, how the Replica benchmark works, what the AI scientist actually beat — and the caveats, because the evaluation was run by Inherent itself. The machine learning details matter here, and so does the training data behind them, so we will separate the verified numbers from the framing throughout.
Table of contents
- What Inherent Announced About Its AI Scientist
- How the Faraday AI Scientist Actually Works
- Inside Replica, the Benchmark Behind the Claim
- The Results: How Faraday Outperformed Claude and Codex
- The Caveats Every AI Scientist Claim Deserves
- The Company Behind Faraday: Funding, Founders and Focus
- What the AI Scientist Race Means for Business
- AI Scientist FAQ
- References
What Inherent Announced About Its AI Scientist
The announcement bundles three things: a model, a benchmark, and a claim about the first competitive AI scientist. Keeping them distinct is the fastest way to understand what has — and has not — been demonstrated.
A 27B-parameter agent called Faraday
Faraday is what Inherent calls an AI scientist: an agent trained to direct the replication of a published paper’s results. It is not a frontier-scale model. It is a 27-billion-parameter system, post-trained from the open Qwen3.6-27B base model using LoRA adapters and long-horizon reinforcement learning. That an agent this size can top far larger systems on any research task is the heart of the story — and the reason the report travelled far beyond the AI research community.
A benchmark called Replica
Alongside the agent, Inherent released Replica, a benchmark of 310 figure-replication tasks drawn from 100 well-known papers in ML and AI-for-science fields, spanning work published between 1990 and 2026. Each task redacts a figure from a paper and asks the agent to reproduce it from the methods described. The domains stretch from NLP and meta-learning to materials science, weather forecasting and structural biology.
The claim itself
In Inherent’s evaluation, the AI scientist outperformed both Claude Opus 4.8 and GPT-5.5 Codex on 73% of in-distribution tasks, and on 60% of a held-out AI-for-science test set — averaging a 6% score improvement over Claude and 8% over Codex on the test split. Those figures come from the arXiv paper, not the TechCrunch piece, which quoted no scores. The headline numbers at a glance:
| Result | Figure | Stated in |
|---|---|---|
| Win rate, in-distribution ML tasks | 73% vs both Claude Opus 4.8 and GPT-5.5 Codex | Inherent’s arXiv paper |
| Win rate, held-out AI-for-science tasks | 60% | Inherent’s arXiv paper |
| Average score gain on the test split | +6% over Claude, +8% over Codex | Inherent’s arXiv paper |
| Faraday’s size | 27B parameters (competitor sizes undisclosed) | Inherent’s paper and blog |
How the Faraday AI Scientist Actually Works
The architecture is the most interesting part of the paper, because the AI scientist does not replace the frontier models — it manages one of them.
A coding agent as a tool
Inherent calls the design “Coding Agent as a Tool”, or CAT. Faraday plays the scientist: it reads the paper, plans the replication, sets the budget and judges intermediate results. The actual code is written by OpenAI’s GPT-5.5 Codex, which Faraday directs the way a principal investigator directs a research engineer. The claimed skill is not programming — it is scientific intuition about what to try, what to trust and when to stop. In effect, the AI scientist supplies the judgement and rents the hands.
Trained by long-horizon reinforcement learning
Faraday’s post-training used long-horizon reinforcement learning — specifically Group-Relative Policy Optimization — over multi-step replication episodes, with LoRA adapters (rank 128) rather than full fine-tuning. The agent never sees the redacted figure it is trying to reproduce; the reward comes from how well its finished replication matches. That makes replication a genuine test of method-following rather than answer-recall, which is precisely the property that makes an AI scientist trainable at all.
Why a small AI scientist can win
The economics of the design deserve a moment. Training a frontier-scale model costs hundreds of millions; post-training a 27B open model with LoRA adapters costs a tiny fraction of that. By spending its training budget purely on scientific judgement — and outsourcing code generation to a commodity coding agent — Inherent concentrated every reinforcement learning update on the one skill the frontier models were never explicitly trained for. A specialised AI scientist does not need to know everything; it needs to know whether the work in front of it is right.
Why replication is the training ground
Replication has two virtues as a task: the ground truth already exists, and success requires understanding a method well enough to re-run it. It is also drudge work that the research community chronically under-rewards — which is why our earlier report on AI research agents taking on original scientific research treated replication as the obvious first rung on the ladder. Inherent has now built an entire training pipeline on that rung.
Inside Replica, the Benchmark Behind the Claim
A result is only as strong as the benchmark it is measured on, and Replica is a new one — authored by the same lab that trained the winner.
310 tasks from 100 papers
Replica’s 310 tasks split into 242 training tasks and 68 held-out test tasks. The held-out set leans on AI-for-science papers — the domains furthest from Faraday’s ML training distribution, and so the fairest measure of generalisation the paper offers. For an AI scientist in training, the split is the curriculum: 242 tasks to learn the craft on, 68 to prove it transfers.
How the replications are judged
Each task carries an auto-generated rubric — created with Claude Opus 4.7 — scoring five dimensions: visual fidelity to the original figure, support for the scientific claim, quality of the experimental implementation, resource budgeting, and scientific integrity. A GPT-5.5 Codex judge scores the outputs with multi-sample aggregation, and Inherent validated the judge against human ratings using rank correlation. It is a careful pipeline — and still one in which the referee, the rubric-writer and one contestant are all frontier lab models.
The benchmark at a glance
| Replica property | Detail |
|---|---|
| Tasks | 310 figure replications (242 train, 68 test) |
| Source papers | 100, published 1990–2026 |
| Domains | NLP, meta-learning, materials science, weather forecasting, structural biology |
| Task format | Redacted figure; agent reproduces it from the paper’s methods |
| Judging | Five-dimension rubrics, LLM judge, validated against human raters |
The Results: How Faraday Outperformed Claude and Codex
The published numbers deserve a close reading, because they are simultaneously impressive and narrower than the headline suggests.
The win rates in context
On the in-distribution ML tasks, the AI scientist won 73% of head-to-head comparisons against both frontier agents. On the held-out AI-for-science set — the harder generalisation test — the win rate fell to 60%, with average score improvements of 6% over Claude Opus 4.8 and 8% over GPT-5.5 Codex. A 60% win rate is a real edge, not a rout: on four tasks in ten, a frontier agent still did the better replication.
David, Goliath, and Goliath’s keyboard
The size asymmetry is what makes the result striking: a 27B-parameter AI scientist out-judging systems whose parameter counts are undisclosed but assumed to be far larger. Yet the asymmetry cuts both ways. Faraday delegates its coding to GPT-5.5 Codex — one of the systems it is being compared against. The fair reading is that a small scientific-judgement layer on top of a frontier coder beats the frontier coder alone; it is a system-versus-model comparison, not a straight fight. Both readings agree on one thing, though: the AI scientist layer added measurable value on top of the strongest coding agent available.
What the founders emphasise
Notably, Inherent itself resists the beat-the-frontier framing. Hughes told TechCrunch the methodology matters more than the results, and described the company’s direction plainly: “We’re always guided by that north star of building an AI scientist agent.” The result, on that reading, is a proof that scientific judgement is trainable — not a leaderboard coup.
The Caveats Every AI Scientist Claim Deserves
Extraordinary benchmark claims deserve the same scrutiny an AI scientist would apply to a paper. Four caveats stand out, all acknowledged in or evident from the sources themselves.
The evaluation is self-run
Inherent authored the benchmark, trained the winning agent, generated the rubrics with one competitor’s model and judged with the other’s. Nothing about that is hidden — it is all in the paper — but no independent replication of the replication result exists yet. Until a third party runs Replica, the figures are Inherent’s figures. That does not make the AI scientist result wrong — it makes it provisional.
The paper flags its own limits
The paper concedes that Faraday “failed in several cases where we have confidence that the original result was obtained rigorously”, and that its human study was too limited to conclude whether people actually prefer Faraday’s replications on average. Both admissions are to Inherent’s credit — and both belong next to the headline.
What is verified, and what is not
| Claim | Status |
|---|---|
| 73% / 60% win rates, +6% / +8% score gains | Stated in Inherent’s arXiv paper; not independently verified |
| Faraday is 27B, built on Qwen3.6-27B with RL | Stated in the paper and Inherent’s research post |
| Faraday made a new scientific discovery | Not claimed anywhere — replication only |
| How much bigger Claude and Codex are | Unknown — competitor parameter counts undisclosed |
Replication is not discovery
Nothing in the release demonstrates new science. Faraday reproduces what human researchers already found — a vital capability, and the stated first step on Inherent’s roadmap, but a different thing from the “genuine scientific discovery” its lead investor describes as the destination. The gap between the two is where the next paper will be judged.
The Company Behind Faraday: Funding, Founders and Focus
The lab making the claim is young, small and unusually well connected — worth knowing, because its structure shapes how much weight to give its roadmap.
From stealth to seed in one spring
Inherent emerged from stealth at the end of May 2026 with a $50 million seed round co-led by Index Ventures and Radical Ventures, joined by NVentures — NVIDIA’s venture arm — plus Ex/Ante, Metaplanet, Macroscopic Ventures and Mythos Ventures. Matt Clifford, the Entrepreneurs First co-founder and former UK government AI adviser, advises the company, which is structured as a public benefit corporation.
A dozen people in King’s Cross
The team is based in King’s Cross, London — a short walk from DeepMind’s home turf — with roughly 12 employees at the time of the TechCrunch report and plans to reach 20 to 25 by year-end, all working in person. Danny Rimer of Index Ventures frames the mission memorably: “Most AI is built to answer questions. What it can’t do yet is figure out which questions are worth asking.”
| Inherent fact file | Detail |
|---|---|
| Founded by | Tantum Collins, Edward Hughes, Louis Kirsch (ex-DeepMind); Kaloyan Aleksiev (infrastructure, ex-Reka AI and Microsoft) |
| Emerged from stealth | Late May 2026 |
| Funding | $50M seed, co-led by Index Ventures and Radical Ventures |
| HQ and team | King’s Cross, London; ~12 people, growing to 20–25 |
| Structure | Public benefit corporation |
| First release | Faraday agent + Replica benchmark, August 2026 |
The AI scientist field is getting crowded
Inherent is far from alone in chasing this goal. Every frontier lab now runs research-agent programmes of its own, and automated-science benchmarks are multiplying — a race we mapped in our earlier report on AI research agents and original scientific research. What distinguishes Inherent’s entry is focus: rather than a general assistant that can also do science, it is training an AI scientist as the product itself, with replication as the measurable first milestone. In a field where claims outrun evidence, picking a checkable milestone is itself a differentiator.
Why the open-model foundation matters
Faraday’s base model is Qwen3.6-27B — an open-weight model, post-trained with a comparatively modest LoRA setup. A $50M-seed startup credibly challenging frontier labs on a research task, using open weights and clever training rather than raw scale, says as much about the state of the open-model economy as it does about Inherent — a dynamic we also traced in our coverage of the reported Hugging Face sale exploration.
What the AI Scientist Race Means for Business
A research-replication agent sounds far from day-to-day operations. Three practical lessons say otherwise.
Small, specialised agents can beat giants
The clearest business signal is architectural. Faraday shows a small model trained for judgement, directing a big model used as a tool, beating the big model used alone. That pattern — specialise the orchestrator, commoditise the executor — applies directly to commercial automation, and it is the approach we take when scoping client systems through our AI strategy service. The lesson holds even if Faraday’s exact numbers soften under independent testing.
Verification is the skill that transfers
An AI scientist is, at bottom, an agent that checks work against a method statement without being shown the answer. Every business process that suffers from unverified outputs — reconciliations, compliance evidence, data pipeline changes — wants exactly that capability. It is the same evaluation gap we examined in our report on AI agent reliability with messy documents: the hard part of agents is not doing the task, it is knowing whether the task was done right.
A London AI scientist lab is a UK signal
There is also a domestic angle. An AI scientist lab of this ambition choosing King’s Cross over San Francisco — as a public benefit corporation, advised by a former UK government AI adviser — is a data point about where frontier-adjacent talent is willing to build. For UK organisations, that proximity is practical: the research culture that produces these agents increasingly sits in the same city as the businesses that will deploy them, and early conversations are easier when the lab is a train ride away rather than a time zone.
Watch benchmarks the way you watch products
Finally, Replica itself is the durable artefact. Benchmarks shape what labs train for, and a public benchmark for replicating research will pull the whole field toward verifiable, method-following behaviour. Teams tracking which capabilities are about to become dependable — via resources like our AI models and tools hub — should treat new benchmarks as a leading indicator, arriving a release cycle ahead of the products.
AI Scientist FAQ
What exactly is Faraday?
Faraday is Inherent’s 27-billion-parameter AI scientist agent, post-trained from the open Qwen3.6-27B model with long-horizon reinforcement learning. It plans and judges the replication of published research papers, delegating the coding to OpenAI’s GPT-5.5 Codex as a tool.
Did it really beat Anthropic’s and OpenAI’s models?
On Inherent’s own benchmark and evaluation, yes: it outperformed Claude Opus 4.8 and GPT-5.5 Codex on 73% of in-distribution tasks and 60% of held-out AI-for-science tasks. The evaluation was designed, run and judged by Inherent, and no independent verification exists yet.
What is the Replica benchmark?
A set of 310 tasks, drawn from 100 ML and AI-for-science papers, in which an agent must reproduce a redacted figure from a paper using only the described methods. Inherent released it alongside the Faraday paper on arXiv on 13 August 2026.
Who is behind Inherent?
Four co-founders — Tantum Collins, Edward Hughes and Louis Kirsch, all formerly of Google DeepMind, plus infrastructure engineer Kaloyan Aleksiev — running a public benefit corporation in King’s Cross, London, backed by a $50M seed co-led by Index Ventures and Radical Ventures.
Has an AI scientist discovered anything new?
No. This release demonstrates replication of existing findings only. Inherent presents replication as training ground for scientific judgement, with genuine discovery as the long-term goal — a goal no lab has verifiably reached yet.
References
Training AI Scientists to Replicate Research — arXiv
Training AI Scientists to Replicate Research — Inherent
London-based AI lab Inherent emerges from stealth with $50M raise — tech.eu
Inherent: Designing for Discovery — Index Ventures
Inherent raises $50M seed to build AI scientists — The Next Web
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.