Inherent says its 27B science agent beat frontier models

TL;DR: London startup Inherent says its new AI agent, Faraday, beat much larger models from Anthropic and OpenAI at reproducing published scientific results. The bigger point is not the benchmark win; it's Inherent's claim that a 27B model plus reinforcement learning can produce better research behavior than brute-force scale.

Benchmark result

Inherent, a London lab founded by Google DeepMind alumni, says Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 on a narrow but relevant task: independently reproducing the findings of published scientific papers without being given the answer up front.

That matters because replication is a real scientific workflow, not a toy benchmark. Hughes framed it as the kind of thing human researchers do early in their careers. If an agent can read a paper, decide which experiments matter, run them, and recover the result, that is closer to useful lab assistance than another coding benchmark or chatbot eval.

Still, the source here is the company itself via TechCrunch. There is no public paper, no external evaluation, and no detailed score breakdown in the material provided. So the claim is interesting, but thin. Treat it as an early signal, not settled fact.

Model efficiency

The standout detail is model size. Inherent says Faraday runs on Qwen 3.6 with 27 billion parameters, while competing against frontier systems from OpenAI and Anthropic that are much larger.

That does not automatically mean Faraday is better overall. It means Inherent may have found a way to make a smaller model behave better on one high-value task. For builders, that is the more useful story. If you can get stronger task performance from a smaller base model through system design and training, your cost structure changes fast.

A 27B system is still large by most startup standards. But compared with frontier-scale proprietary models, it is a far more plausible foundation for specialized agents where inference cost, latency, and iteration speed matter.

Research behavior

Inherent's actual pitch is not "we beat OpenAI." It is "we taught an agent better research behavior."

Hughes says the company set a higher bar than simple accuracy. It wanted Faraday to show "research taste": choosing worthwhile experiments, structuring them well, and following curiosity rather than just pattern-matching toward an answer.

That is a useful distinction. Many current agents are good at producing plausible next steps and bad at deciding which next steps are worth the compute. In research workflows, that gap is everything. Bad taste means wasted runs, shallow conclusions, and high-confidence noise.

The most credible part of Inherent's thesis is that science agents need judgment, not just tool use. Whether Faraday has that in a robust way is still unproven publicly. But the target is right.

Reinforcement learning stack

Inherent says reinforcement learning is the main mechanism for teaching this behavior. Instead of trying to encode scientific method directly or training mostly on examples of research process, the company rewards good outcomes and bets the resulting policy will generalize across domains.

That approach faces two real risks. First, RL is sample-hungry, and the company has not published its compute spend or training data composition, so reproducibility is open. Second, behavioral rewards can overfit to the specific reward model and collapse into a narrow solution that looks like research taste but is actually reward hacking. The field has seen this pattern before in code-generation RL.

Faraday's hybrid stack — a 27B Qwen base model wired to OpenAI's GPT-5.5 Codex for tool and code execution — is also worth flagging. The "small open-weight brain, large closed model hands" pattern is becoming common in agent work. It splits cost and explainability across two vendors in ways that are easy to underestimate when you read the headline.

What this means for builders

The pattern to watch is not "27B beat 405B." It is "specialized RL-on-behavior on a smaller base can beat general scaling on narrow high-value tasks." That is a real claim with real precedent. Math and code RL-tuned open models have closed gaps with frontier systems over the past year. If Faraday's replication result holds up under independent testing, scientific research assistants are next on the list of agent categories where small plus RL beats large plus general.

For anyone building agents today: do not assume frontier-scale is the only path. A focused training loop on the behavior you actually need — not the behavior the leaderboard rewards — is still underpriced relative to its results.

The honest caveat: one company's blog post and a TechCrunch writeup are not evidence. Inherent needs to publish a method paper, disclose its evaluation set, and let outside labs try to reproduce the result. Until then, treat this as a signal worth tracking, not a verdict.