Inherent Labs Releases 27B Model Faraday, Outperforms Claude Opus 4.8 in Paper Reproduction

Inherent Labs released the 27B parameter model Faraday, positioned as an "AI Scientist." The model employs the Replica method and is trained via long-horizon reinforcement learning to combine agent capabilities with scientific intuition. Official evaluations claim that it outperforms Claude Opus 4.8 and GPT-5.5 on held-out research reproduction tasks. Faraday's core strategy decouples "research direction decision" from "code implementation," taking charge of high-level command while invoking coding tools like GPT-5.5 Codex to complete specific experiments. If these results hold, it demonstrates that specific training paradigms can enable smaller models to surpass flagship models in scientific tasks, showcasing the feasibility of small models directing large coding models for research.

Confirmed

Inherent Labs released the 27B parameter model Faraday, trained via long-horizon reinforcement learning using the Replica method.
The task involves removing a result chart from a paper and requiring the agent to reproduce the result within limited compute.
The model decouples research planning from code implementation, directing tools like GPT-5.5 Codex to conduct experiments.
Official claims state it defeats Claude Opus 4.8 and GPT-5.5 on held-out research reproduction tasks.
Target applications include climate science, structural biology, and materials science.

Unconfirmed

Comparison results are from the developer's self-evaluation; independent third-party verification is pending.
Details on evaluation benchmarks, training data composition, and whether weights or APIs are open were not disclosed.
dhadfieldmenell pointed out limitations regarding paper reproduction capabilities; specific details of the criticism are unknown.

2026-08-14 ~ 2026-08-16 · 16 related posts

Primary sources

27B Research Agent Outperforms Opus and GPT-5.5 — rohanpaul_ai · 2026-08-15 07:18 (UTC+08:00)
27B agent Faraday beats Claude Opus 4.8 and GPT-5.5 on research replication via new Replica method — omarsar0 · 2026-08-15 01:00 (UTC+08:00)
27B Faraday model directs GPT-5.5 to reproduce papers, beats Claude Opus 4.8 — aigclink · 2026-08-15 12:52 (UTC+08:00)

Faraday: 27B AI Scientist Outperforms Claude Opus 4.8 on Paper Replication — Tom_Westgarth15 · 2026-08-14
Inherent Releases First Public AI Scientist: Faraday — Scobleizer · 2026-08-15
[source] 27B agent Faraday beats Claude Opus 4.8 and GPT-5.5 on research replication via new Replica method — omarsar0 · 2026-08-15
Inherent Releases Faraday: AI Scientist Outperforms GPT-5.5 in Replicating Papers — ricklamers · 2026-08-15
[source] 27B Research Agent Outperforms Opus and GPT-5.5 — rohanpaul_ai · 2026-08-15
[source] 27B Faraday model directs GPT-5.5 to reproduce papers, beats Claude Opus 4.8 — aigclink · 2026-08-15

10 near-duplicate retellings: matthen2 · ChengleiSi · YesThisIsLion · aigclink · gleech · dhruv2038 · jzl86 · burny_tech · ChenhaoTan · neal_lathia