AI Daily — 2026-07-24

English 中文

Claude Opus 5 debuts, near Fable 5 at half price, with NVIDIA championing open models and Huang debating open vs closed AI distillation.

Covering 18 AI news items

🔥 Top Stories

1. Claude Opus 5 Debuts, Nears Fable 5 at Half Price

Anthropic unveils Claude Opus 5 as a thoughtful, proactive model that allegedly approaches frontier capabilities at about half the price of peers like Fable 5. If early benchmarks hold, Opus 5 could sharpen price-performance competition and broaden access to top-tier AI. The move also intensifies safety and governance debates as buyers weigh capability against risk. Source-x

2. NVIDIA signs letter on why open models matter

NVIDIA publishes a signed letter arguing open models are essential for safety, cybersecurity, rapid innovation, and broad-based AI development. It contends AI will transform every industry and be built by multiple nations, necessitating a mix of frontier open and frontier closed models. A linked PDF provides deeper rationale and policy context. Source-x

3. Huang on AI Distillation: Open Source vs Closed Models

Jensen Huang discusses distillation in an Axios interview, arguing AI must continually learn from humans, other AI, and diverse sources of knowledge. He notes the internet could become largely AI-generated, implying AI systems will distill knowledge from other AIs. He also weighs whether open-source model makers should distill from closed models, highlighting ongoing industry debates. Source-x

Generated by AI News Agent | 2026-07-24

📰 Featured

LLMs & Benchmarking

Fugu-Ultra v1.1 Released with Stronger Benchmarks and Capabilities — Upgrades to the Fugu family yield stronger benchmarks (notably ProgramBench and Terminal Bench 2.1) and improved coding/agentic abilities, while price remains unchanged. Source-x
GPT-5.6 Pro Touted Top Model for Hard Tasks — Claims GPT-5.6 Pro is the smartest for hard tasks, ranking above Fable 5 Ultracode and GPT-5.6 Sol Ultra, though the lineup is described as confusing. Source-x

Hardware & Training

SLAI T-Rex: Full-Parameter Post-Training on Ascend SuperPOD — Full-parameter MoE training reveals system bottlenecks (memory pressure, non-overlapped communication); hardware-aware optimizations on Ascend with the DeepSeek-V4 workload illustrate feasibility at scale. Source-huggingface

Embodied AI & Multimodal

ReferTrack Enables Referring-Then-Tracking for EVT — Proposes resolving target references before tracking to align reasoning with explicit detections, addressing limitations of vision-language-action policies that rely on latent representations. Source-huggingface

Open Source Tools & Productivity

ego-lite browser enables parallel browsing for you and AI agents — A browser designed for simultaneous use by humans and AI agents, with agents in separate Spaces while your tabs stay private; macOS available today with Windows/Linux roadmap; GitHub demo and install steps. Source-github

Cheap, Fast Tool-Use Models

AntLing-3.0-flash: cheap fast AI model for tool use — 124B model focused on tool-use execution rather than heavy reasoning; API-only on OpenRouter with free access through Aug 3, signaling a trend in affordable, fast tool-use models from Chinese labs. Source-reddit

Open Weighting & Safety Advocacy

American Tech Company Advocates Open Weighting Models in Letter — Open weighting models are advocated as essential for a healthy AI ecosystem; warns against exclusive reliance on closed AI models. Source-reddit

⚡ Quick Bites

Anthropic: stop Claude persona, focus on context window — Source-x
AREX: Recursively Self-Improving Agent for Deep Research — Source-huggingface
Removed 80% of Claude Code prompts; lessons learned — Source-x
K12-KGraph Establishes Curriculum-Aligned Knowledge Graph for Educational LLMs — Source-huggingface
Visual Contrastive Self-Distillation Removes Asymmetry via Input Conditioning — Source-huggingface
Earthtojake text-to-cad launches AI CAD agent skills library — Source-github
Harper: Offline, privacy-first Rust grammar checker by Automattic — Source-github
DeepSeek CEO: 6x Profit Restraint, Open-Source Policy — Source-reddit