🤖 AI Daily · Jul 25, 2026

Anthropic launches Claude Opus 5: near-Fable intelligence at half price; Ant Group's Ling-3.0-flash: 124B params, only 5.1B activated for hybrid reasoning; Black Forest Labs releases FLUX 3 multimodal model supporting 20-sec video with native audio; Midjourney V8.2 focuses on aesthetic quality; Runway Agent introduces natural language workflow builder; Claude Code v2.1.219 adds Opus 5 with 1M context; NVIDIA, Microsoft, Meta jointly warn against over-regulating open-weight AI models; Kimi K3 trails US frontier models on cyber exploit tests; Anthropic launches Drone-Bench for AI drone control evaluation; Apple proposes LEAD to solve long-horizon reasoning bottlenecks.

1. 1. Anthropic Launches Claude Opus 5: Near-Fable Intelligence at Half Price

   Claude Opus 5 offers near-Fable 5 intelligence at half the price. Doubles Opus 4.8 on Frontier-Bench, triples runner-up on ARC-AGI 3. Default model for Claude Max.

2. 2. Ant Ling-3.0-flash: 124B Params, Only 5.1B Activated

   Ant Group's Ling-3.0-flash hybrid reasoning model uses 124B total params with only 5.1B activated, surpassing predecessor Ring-2.6-1T with linear attention and 1/64 sparse MoE.

3. 3. Black Forest Labs Releases FLUX 3: 20-Second Video with Native Audio

   FLUX 3 jointly trains image, video, and audio, generating up to 20-second videos with native audio in one pass. Partners with mimic for Audi production line testing.

4. 4. Midjourney V8.2: Aesthetic Quality Focus

   Midjourney V8.2 enhances aesthetic quality and personalization, significantly reducing low-quality outputs with better understanding of user style preferences.

5. 5. Runway Agent Launches Natural Language Workflow Builder

   Runway Agent now supports building, running, and editing node-based workflows via natural language, unlocking high-quality video output at scale.

6. 6. NVIDIA, Microsoft, Meta Warn Against Over-Regulating Open-Weight Models

   NVIDIA, Microsoft, and Meta sign open letter warning over-regulation of open-weight models would hurt US AI competitiveness. OpenAI and Anthropic did not sign.

7. 7. Kimi K3 Trails US Frontier Models on Cyber Security Tests

   UK-US joint evaluation: Kimi K3 scores 32.2% on ExploitBench vs 76.2% for US frontier models, but ahead of GLM-5.2 at 24.4%. Knowledge distillation may explain gap.

8. 8. Anthropic Launches Drone-Bench for AI Drone Control Evaluation

   Anthropic and Andon Labs introduce Drone-Bench, testing AI models' ability to autonomously pilot quadcopter drones for indoor person定位 and tracking.

9. 9. Apple Proposes LEAD to Solve Long-Horizon Reasoning Bottlenecks

   Apple research identifies 'no-recovery bottlenecks' in LLM long-horizon tasks. LEAD method breaks error accumulation via short-horizon future verification and aggregation.

10. 10. Claude 5th Gen Context Engineering: System Prompts Reduced by 80%+

    Anthropic removes 80%+ of Claude Code system prompts for Opus 5 and Fable 5 without significant coding benchmark degradation, redefining context engineering.