Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-24 · high
Inherent says reinforcement-learned research taste, not scale, drove the win, and the DeepMind spinout raised a $50 million seed.
Why now: TechCrunch reported the claim and the funding on August 22, 2026.
Inherent, a British lab founded by DeepMind alumni, says its agent Faraday reproduced published research findings without seeing the results in advance, and did it better than Anthropic's Opus 4.8 and OpenAI's GPT-5.5. Faraday runs on Qwen 3.6 at 27 billion parameters, tiny next to frontier models that run into the hundreds of billions.
The team's explanation is reinforcement-learned research taste. Instead of scaling parameters, they trained the model to pick which experiments to run, according to TechCrunch's report. If that holds up, it challenges the assumption that better research agents need bigger base models.
Inherent raised a $50 million seed alongside the claim. That's real money betting a smaller, taste-tuned model can out-navigate paper reproduction, a hard task: pick the right experiment, run it cleanly, and land on the number the original authors got.
TechCrunch doesn't publish comparative scores, so there's no way to check the margin or the benchmark it was measured on. Reproducing paper results rewards a model that's good at picking experiments over one that's merely bigger, so the mechanism is plausible. But "outperformed Anthropic and OpenAI" without numbers is Inherent's framing of its own eval, unverified.
If you're building research agents, watch whether Inherent publishes the benchmark and the scores behind this claim. Until then, it's unquantified.
Each link below shares sources, entities, or timing with this story.
DeepMind competes with OpenAI / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (DeepMind competes with OpenAI); both cover GPT, Opus, TechCrunch; reported by the same outlet (techcrunch.com).
Linked by a graph relationship (DeepMind competes with OpenAI); both cover GPT, Opus, TechCrunch; reported by the same outlet (techcrunch.com).
DeepMind competes with Anthropic / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (DeepMind competes with Anthropic); both cover GPT, Opus, Qwen; earlier GPT coverage from 2026-04-21.
DeepMind competes with Anthropic / Shared entities / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (DeepMind competes with Anthropic); both cover Opus, TechCrunch; reported by the same outlet (techcrunch.com).
DeepMind competes with OpenAI / Shared entities / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (DeepMind competes with OpenAI); both cover Opus, TechCrunch; reported by the same outlet (techcrunch.com).
Claude Code uses Opus / Shared entities / Same source domain / Earlier coverage / Downstream implication
Linked by a graph relationship (Claude Code uses Opus); both cover Opus, TechCrunch; reported by the same outlet (techcrunch.com).
DeepMind competes with OpenAI / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (DeepMind competes with OpenAI); both cover GPT, TechCrunch; reported by the same outlet (techcrunch.com).
DeepMind competes with Anthropic / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (DeepMind competes with Anthropic); both cover Opus, TechCrunch; reported by the same outlet (techcrunch.com).