A 27B Model Just Outperformed GPT-5.5 at Science
Inherent, founded by DeepMind researchers, just released Faraday—a compact AI agent that beats larger models from OpenAI and Anthropic at the specific task of replicating published research papers.
A Smaller AI Just Beat the Frontier Models at a Specific Game
<cite index="1-1">On August 14, 2026, London-based Inherent introduced Faraday, a 27-billion-parameter "AI Scientist" agent designed to replicate scientific research</cite>. The interesting part? <cite index="2-3,2-4">Faraday, built on the relatively small 27-billion-parameter Qwen 3.6 model, was tested against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5—both significantly larger frontier-scale systems—and achieved higher success in replicating research results without being given the answer in advance</cite>.
This might sound like a narrow claim. It is. But it's also pointing at something real about how AI agents might work in the near future.
What Actually Gets Tested Here?
<cite index="1-6,1-7">Inherent introduced Replica, a scalable set of 310 tasks drawn from 100 machine-learning and AI-for-science papers spanning natural-language processing, materials science and weather forecasting, where each task asks an agent to reproduce a figure from a paper without access to the original plot, under limited time and compute budgets</cite>.
The catch: <cite index="1-8">the benchmark is intended to test the experimental reasoning and persistence needed to recover the "99% perspiration" that research papers usually omit, rather than simply reproduce a known answer</cite>. This isn't about plugging numbers into a formula. It's about inference, iteration, and what the team calls "research taste."
The Architecture That Actually Matters
<cite index="1-13">Faraday uses GPT-5.5 Codex as a tool, directing a model several orders of magnitude larger in a way that Inherent says improves replication performance</cite>. In other words: Faraday doesn't replace GPT-5.5's coding ability. It sits on top of it, trained to ask smarter questions and pursue better hypotheses.
<cite index="19-4,19-5,19-6">Faraday is trained with long-horizon reinforcement learning, using coding agents as tools</cite>. The training process used <cite index="22-5">an auto-generated rubric-based judge that has low noise and agrees with human assessment of replication quality</cite>.
The Numbers (Where Faraday Actually Shines)
<cite index="6-11,6-12">On machine learning tasks similar to its training data, Faraday outperformed Claude on 73% of tasks. On a held-out test set of "AI-for-science" papers (e.g., physics, biology), Faraday still held a lead, outperforming baselines on 60% of tasks</cite>.
Most revealing: <cite index="1-11">the company says Faraday also struggles less with recent research, suggesting stronger out-of-distribution performance on work unseen during the base model's pre-training</cite>. Newer papers are harder for most models because they use complex distributed systems and massive compute—exactly the kind of underspecified problem Faraday was trained to handle.
Why This Matters for Developers
<cite index="13-2">The company was founded by former DeepMind researchers Tantum Collins, Edward Hughes, and Louis Kirsch, joined by Kaloyan Aleksiev (formerly at Reka AI and Microsoft)</cite>. <cite index="1-4">The launch follows Inherent's emergence from stealth and its reported $50M seed round led by Index Ventures, with Radical Ventures also participating</cite>.
The real insight: this work suggests that smaller, specialized agents trained with RL to handle underspecified, exploratory problems might outperform raw scale. That's a useful pattern as the AI space matures. <cite index="22-8">The researchers believe that their results provide a stepping stone towards AI agents capable of long-horizon scientific innovation without requiring complex harnesses</cite>.
For your own workflows, this hints at an emerging approach: rather than asking a single large model to do everything, train or fine-tune smaller agents to excel at specific reasoning patterns (like scientific replication), then compose them with larger models. If you're building research tools or scientific workflows, tools like NeonCodex AI can help you orchestrate these kinds of agent combinations at scale.
The Caveat
<cite index="11-8">This may sound like a mere party trick given Inherent's much loftier goal — building AI that can discover new scientific knowledge and not just verify old results</cite>. Replication is a stepping stone, not the destination. But stepping stones matter. They show what's possible.
Inherent isn't open-sourcing Faraday or making it available as a public API yet. <cite index="11-11">The startup plans to grow its headcount to about 20 to 25 by the end of the year</cite>, so the team is likely still focused on proving the model's value for internal discovery work.
What to Watch
The architecture here—a smaller agent trained on underspecified reasoning tasks, leveraging a larger model's coding ability—is worth studying if you're building agentic AI systems. <cite index="20-5,20-6">Faraday's improvement was not just about better coding; when researchers tried to "prompt-engineer" the baseline models to act more like scientists, the gap remained</cite>. That tells you something important: structure and training matter more than you might expect.
If you're working with agents in production, start running tests on whether fine-tuned smaller models with long-horizon RL can outperform your current setup on specific, underspecified tasks. You might find the same efficiency gains Inherent did.
Source: [TechCrunch](https://techcrunch.com/2026/08/22/inherent-founded-by-deepmind-alumni-says-its-ai-teammate-just-outperformed-anthropic-and-openai-at-replicating-research/)