Section

Models

Model releases, capability changes, benchmarks, and deprecations.

4 published stories

TechCrunchAug 23, 2026Models

Who’s behind the new ‘stealth model’ Ox Alpha?

Speculation has emerged surrounding a newly surfaced stealth artificial intelligence model named Ox Alpha, following its appearance in developer and benchmarking circles such as OpenRouter. The identity of the underlying developer, architectural specifications, deployment timeline, and technical benchmarks remain unconfirmed publicly. Industry observers are analyzing early telemetry and output characteristics to determine whether the model originates from an established frontier AI lab or an emerging independent research team.

28/100

TechCrunchAug 22, 2026Research

Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

British startup Inherent, founded by former Google DeepMind researchers, has introduced Faraday, an AI agent designed to replicate scientific research papers. The company claims the tool outperforms frontier models from OpenAI and Anthropic in scientific replication workflows. The system is positioned as an AI collaborator to accelerate scientific discovery and validate published findings, though the comparative performance metrics currently reflect vendor-reported evaluations rather than independent peer review.

34/100

TechCrunchAug 21, 2026Security

Anthropic’s Opus 4.6 is a smut-machine

Independent testing by TechCrunch revealed that Anthropic's Claude Opus 4.6 model can be readily prompted to bypass built-in safety guardrails prohibiting sexually explicit material. Despite Anthropic's stated usage policies restricting adult content generation, reporters demonstrated that standard jailbreaking techniques successfully circumvented the model's automated moderation filters during testing.

49/100

TechCrunchAug 21, 2026Research

Nvidia just showed that the harness, not the AI model, is now the real hero

Nvidia published research demonstrating that structured agent harnesses and targeted fine-tuning enable AI agents to perform reliably, even when powered by less capable underlying foundation models. The study highlights that execution scaffolding, guardrails, and task-specific tuning effectively prevent agent drift and hallucination during multi-step tasks. This approach demonstrates that engineering the surrounding software architecture can compensate for limitations in baseline model size and capability.

51/100