Inherent Says Its 27B Agent Beat Claude Opus and GPT-5.5

A London lab says its agent beat Claude Opus 4.8 and GPT-5.5 at reproducing scientific papers while running on a 27-billion-parameter model.

Inherent, founded by Google DeepMind alumni, released the agent — called Faraday — weeks after emerging from stealth with a $50 million seed round. TechCrunch reported the claim on August 22. The same week, a separate benchmark of Claude Code against Vercel's eve framework pointed at the same conclusion from a different direction: the scaffolding around the model is now moving the numbers more than the model tier is.

What Inherent claims

The task was narrow and checkable: "independently reproducing the findings of published scientific papers without being told the answer in advance." Inherent measured Faraday against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, described by TechCrunch as "both much larger, frontier-scale systems." Faraday runs on Qwen 3.6, a model with 27 billion parameters — parameters being, as the report puts it, "a proxy for a model's size and, typically, its training costs, as well."

Two caveats before anyone reorganizes a stack around this. First, the comparison is Inherent's own, reported as the company's claim rather than as an independent evaluation, and the report carries no per-model scores. Second, paper replication is one task. Cofounder and chief scientist Edward Hughes framed it as a starting exercise rather than a summit: "Many PhD students actually start by doing this."

Hughes was also explicit that the ranking was not the interesting part. "What was most interesting to us about this was not so much the result of beating those frontier agents — which of course we liked — but was actually the way we went about building this."

The part that isn't about model size

Inherent's stated bar went past accuracy. It wanted Faraday to show what the company calls "research taste" — "an instinct for what experiments are worth running and how to design them well." Its approach to teaching that is reinforcement learning, described in the report as "a training method that rewards an AI system for good outcomes rather than spelling out rules for it to follow," chosen on a bet that reward-based training generalizes across scientific fields better than training on how science is conducted.

The most concrete signal for anyone building agents is what Inherent chose not to build. Rather than developing its own coding tool, the company had Faraday use OpenAI's GPT-5.5 Codex, "much the way human scientists lean on existing software rather than building everything themselves, according to the company." A small model, someone else's coding tool, and the team's own training loop and scaffolding — that is the whole stack described.

Company shape, for scale: a dozen employees working in person out of King's Cross in London, with plans to reach "about 20 to 25" by the end of the year.

Same models, different harness

The second data point is cleaner because it holds the model fixed. Ora, a platform that benchmarks AI agents on live websites, ran Claude Code against Vercel's eve framework "across hundreds of real user journeys." Both harnesses ran the same models — Claude Fable 5 and Haiku 4.5 — and per the write-up, "the only variable was the harness—the software that gives the model tools and drives it step by step." eve came out ahead on every axis reported, including 7% fewer steps to reach the goal and twice the native success rate, meaning it finished on the customer's own site instead of falling back to web search. We covered those figures when they landed; the reason to raise them again here is the structure of the result, not the percentages.

Two details from that write-up frame the difficulty. Ora "estimates 99% of the web isn't agent-ready," which is the floor these numbers were measured against. And Ora reportedly found a prompt-caching issue in eve "that led to a 15% cost reduction after fixing" — a plumbing bug, not a model limitation, worth 15% of the bill.

Scope matters too. As the write-up puts it, "Claude Code is a coding agent. It excels at repo-level tasks: editing files, running tests, committing code." The benchmark tested something else entirely — navigating live sites, filling forms, calling APIs. A harness losing at a task it was not built for is not the same as a harness being worse.

Both numbers come from interested parties

This is the load-bearing caveat. The Faraday comparison is Inherent's account of its own agent's performance. The Ora benchmark results were, per the write-up, "published on Vercel's blog" — that is, a benchmark favorable to eve appearing on the site of the company that builds eve. Neither is fabricated, and neither is a neutral evaluation. Treat both as vendor-reported until someone reruns them.

What survives the discount is the direction of travel, because the two claims are structurally independent. One says a 27B model plus good training and someone else's coding tool competed with frontier-scale systems on a specific task. The other says an identical model pair produced measurably different outcomes under two harnesses. Neither requires the other to be true.

What to do with this

If you are choosing an agent stack this quarter, the practical read is that harness and model are separate purchases and you probably have not priced them separately. Before your next model upgrade, check three things the reports above found mattered: whether wrong tool calls surface as errors or as plausible-looking results, whether your agent falls back to web search when the site's own endpoints would work, and whether your prompt caching is actually working — that last one was worth 15% in eve's case.

What to watch: whether anyone outside Inherent reproduces the Faraday result with published per-model scores, and whether Ora's journey benchmark gets run by someone who does not sell one of the harnesses. Until then, the strongest version of this week's news is not "small models won." It is that two separate teams found the expensive variable was not the model.