DeepMind AI Startup Claims It Beat OpenAI and Anthropic

A London AI lab, financed by a $50 million seed round and staffed by a dozen former DeepMind researchers, claims its prototype agent has just beaten the biggest models from OpenAI and Anthropic at a core scientific task. According to TechCrunch, Inherent's agent, named Faraday, outperformed GPT-5.5 and Claude 4.8 Opus at independently reproducing the findings of published research papers. The stunner: Faraday does this while running on a model with just 27 billion parameters, a fraction the size of the trillion-parameter "frontier" models it's compared against.

This is not another incremental chatbot update, echoing broader concerns about the shift from static chatbots to active AI agents that can act autonomously, a trend recently underscored by OpenAI's own product chief declaring the chatbot era over and the confirmation that AI hackers are now a routine threat alongside the proliferation of AI swarms targeting critical APIs and moves by companies like Gamma to create human-feeling presentations. It's a direct challenge to the dominant AI narrative that equates capability with scale, a philosophy that continues to face legal scrutiny regarding its data sources, as seen in a recent judge's fair-use AI training ruling, and that is similarly being challenged by companies aiming to automate AI development roles. Inherent's founding team, steeped in the agent-based, goal-oriented AI that defined DeepMind's greatest hits, is betting that the future isn't a smarter conversationalist, but a capable, autonomous colleague. Their benchmark result suggests that for complex, real-world tasks, specialized intelligence trained for action may already be outpacing generalized intelligence trained for conversation.

The Quiet Assault on AI's Self-Perpetuating Laundry Cycle

The core narrative from major AI labs has been one of escalating scale: more parameters, more data, more compute, leading to a more capable general-purpose assistant. The product is often a refined chat interface. Inherent's Faraday represents a stark divergence. It's built not to converse, but to act.

Its target is the "messy interface of the real world," specifically the experimental workflow of a research lab. Where prompting a frontier model to summarize a paper tests comprehension, Faraday's benchmark tests execution: reading a paper's methods, formulating a plan, writing code to run simulations or analyze data, and reproducing the results without being given the answer. This frames AI's next key battle not on the usability of a chat window, but on its competence in operational environments. It’s a shift from building a genius librarian to training a competent lab technician who can work unsupervised.

"What was most interesting to us about this was not so much the result of beating those frontier agents... but was actually the way we went about building this," said Inherent cofounder and chief scientist Edward Hughes.

The "way they built it" is a direct rebuttal to the scale-at-all-costs mentality. It suggests a plateau for generalist models on certain complex tasks, one that specialized, efficiently-trained agents can leap over.

Faraday's 27-Billion-Parameter Jump from Chat to Action

Operationally, Faraday is not a prompted chatbot. It is an AI agent, a system that perceives its environment (a research paper and a digital lab workspace), takes actions (designing experiments, writing code), and aims to achieve a defined goal (successful replication). This is a fundamentally different architecture from using an API call to GPT-5.5 to summarize text.

The replication benchmark is critical. It's not simple recall or summarization; it requires parsing mathematical models, understanding methodological constraints, translating described procedures into executable code, and interpreting results. It's a holistic test of reasoning, planning, and tool use.

Performance isn't just accuracy; it's "research taste." Inherent wanted Faraday to demonstrate an instinct for what experiments are worth running and how to design them well. They cultivated this not by feeding it more scientific papers, but by using reinforcement learning (RL) to reward the agent for good outcomes and sound methodology. This is a page taken directly from the DeepMind playbook used to master games like Go and StarCraft.

Faraday’s pragmatic architecture highlights its focus on the end goal, not ownership of every component. Rather than build its own coding tool, Faraday uses OpenAI's GPT-5.5 Codex, "much the way human scientists lean on existing software." This efficiency-first approach allows the small team to focus its innovation on the agent's core reasoning and planning layers.

A 99% Smaller Model That Outperforms on a $50 Million Bet

The economic implications of Faraday's architecture are as significant as its performance. A 27-billion-parameter model requires orders of magnitude less compute for training and inference than a trillion-parameter frontier model. This translates to dramatically lower costs and latency, making sophisticated AI assistance potentially viable for academic labs and R&D departments without massive budgets.

Inherent's $50 million seed round is a high-conviction bet on this efficiency paradigm, but the runway is finite. Building a reliable "teammate" involves longer, more uncertain development cycles than iterating on a conversational model. The venture capital question is whether backers have the patience for the agentic path when the chat-based path has clearer, near-term revenue models. However, as we've seen in cases like AT&T's 56% AI spending cut by ditching ChatGPT, enterprise appetite for cost-effective, specialized AI is intense and growing.

DeepMind’s Progeny: From AlphaGo’s Intuition to Faraday’s Taste

To understand Inherent's philosophy, look at its lineage. The founding team are Google DeepMind alumni. DeepMind's legacy is defined by agent-based systems that learn to achieve goals: AlphaGo developed intuitive "taste" for board positions, AlphaFold solved the protein-folding problem. These were not large language models; they were specialized agents honed with reinforcement learning.

Faraday is a direct conceptual descendant. It applies the same principles, goal orientation, reward-based learning, strategic planning, to the domain of scientific inquiry. This contrasts sharply with OpenAI's evolution, which pivoted from game-playing agents (like those in Dota 2) to the large language model primacy of the GPT series.

This isn't a minor feature difference. It's a philosophical fork in AI development: one road leads to increasingly capable generalist oracles, the other to a constellation of efficient, specialized digital workers. Inherent is betting the latter will be more transformative for fields like science, where action in the real world is the ultimate metric.

The Tangible Promise: Reshaping the Lab and the Pace of Discovery

If Faraday evolves from a benchmark prototype to a daily tool, its impact could be profound.

For academic research, a functioning AI teammate could accelerate the grind of literature review, hypothesis testing, and replication. It could act as a tireless junior researcher, potentially alleviating routine burdens but also raising existential questions about the training and role of PhD students. The pace of discovery in fields from pharmacology to materials science could increase, but so could the volume of AI-generated results that require new forms of validation.

For industry R&D, the economic case is compelling. A cost-effective agent that can navigate proprietary data, run simulations, and design experiments could compress development cycles and lower barriers to innovation. The promise is not just automation, but augmentation: Hughes models Faraday on his "favorite kind of teammate" who autonomously explores a curiosity and returns with data, asking "What do you think of these results?"

This collaborative vision also brings peril. The ethical and validation crises are immediate: How do we audit an AI's experimental process? Who is responsible for a flawed AI-generated finding? The scientific method is built on human accountability and reproducibility. Integrating an opaque AI agent as a co-author on the process demands new frameworks, a challenge the industry is only beginning to face, as seen in the safety debates that can slow progress, as with OpenAI's reported safety pauses.

2027's Litmus Test: From Simulated Benchmark to a Stainless-Steel Lab

Inherent's next year will determine if Faraday is a fleeting benchmark champion or the progenitor of a new AI category.

In the near term, expect a flurry of "agent" announcements from other labs, each touting superiority on specialized tasks. A war of narrow benchmarks will commence. Inherent plans to grow from 12 to "about 20 to 25" employees by year's end, a hiring push that could attract more DeepMind talent seeking focused, agent-oriented work.

The mid-term, make-or-break test is adoption in a real, non-simulated laboratory. Can Faraday move from replicating known papers in a digital sandbox to assisting with ongoing, novel research? Can it handle the unpredictable noise and physical constraints of real experiments? This transition from demo to daily driver is where most ambitious AI projects falter.

If it clears that hurdle, the long-term implication is a redefinition of the scientist's role and the creation of a new layer of AI infrastructure. This layer would be separate from the chat bots, a platform for digital labor specialized in analysis, design, and discovery. Success would validate a path where AI prowess is measured not by parameter count, but by practical utility, echoing the kind of specialized, high-impact performance that recently shocked mathematicians in another domain.

For now, Inherent has delivered a potent proof-of-concept: that in the high-stakes game of AI, a small, focused team with a deep philosophical blueprint can, in one specific but meaningful arena, checkmate the giants. The game has just gotten more interesting.

Why This Changes Everything

• It challenges the industry dogma that AI capability is solely determined by model size and compute scale, potentially resetting the competitive landscape.
• It demonstrates that specialized, action-oriented 'AI teammates' for complex professional workflows can outperform larger, general-purpose conversational models at specific core tasks.
• It validates a $50M+ investment thesis in agent-based AI, signaling a shift in venture capital and research focus from pure scale to targeted intelligence.