Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-17 · high
Trained on 310 paper-replication tasks scored by an LLM judge, the 27B model also claims an edge over Opus 4.8.
Why now: Inherent published Faraday and the Replica environment on August 14, three days before this August 17 coverage.
A 27B model named Faraday beat Opus 4.8 and GPT-5.5 at replicating research papers, Inherent Labs says. Faraday is a fraction of the size of both frontier models. Beating them on a real task, even a narrow one, is a direct counter-example to the assumption that scale always wins.
Faraday came out of Inherent Laboratories on August 14. It doesn't answer questions or chat. It uses coding agents as tools to recreate figures from published papers, inside fixed time and compute budgets.
The training environment behind it, Replica, pulls 310 reinforcement-learning tasks from 100 papers. Each task gets scored by an LLM judge running a rubric generated for that specific task, checked against human studies, per Inherent's research page. The company says its biggest gains show up in meta-learning, structural biology, and materials science, not general capability.
That edge holds only inside Replica's own rubric, and Inherent wrote both the tasks and the judge that scored them. The result isn't fake, but one test hasn't run: a domain Inherent didn't build the eval for. Until someone runs it, the claim is proven only inside Replica.
Each link below shares sources, entities, or timing with this story.
LLM uses OpenAI / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (LLM uses OpenAI); both cover GPT, LLM; overlapping topics (against, agent, model).
LLM uses OpenAI / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (LLM uses OpenAI); both cover GPT, LLM, Opus; earlier GPT coverage from 2026-04-20.
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Opus built by Anthropic); both cover GPT, LLM, Opus; overlapping topics (agent, model).
Simon Willison released LLM / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Simon Willison released LLM); both cover GPT, Opus; overlapping topics (model, task).
Simon Willison released LLM / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison released LLM); both cover August, GPT; overlapping topics (against, agent, model).
Linked by a graph relationship (Simon Willison released LLM); both cover GPT, LLM; overlapping topics (agent, model, task).
Claude Code uses Opus / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Code uses Opus); both cover GPT, Opus; overlapping topics (beating, model, task).
Cursor uses Opus / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Cursor uses Opus); both cover GPT, Opus; overlapping topics (budget, model, task).