🤖 AI Daily · Jul 18, 2026

Sora 2 achieves indistinguishable deep video cloning; Kimi K3 tops Frontend Code Arena leaderboard; NVIDIA releases Nemotron 3 Embed series; Tongyi's Wan-Streamer v0.2 hits 550ms end-to-end latency; Apple sues OpenAI, legal battle escalates; Moonshot AI reveals Kimi K2.5 technical路線 at GTC 2026; Schema Harness scores 99% on ARC-AGI-3 public set; Claude Fable 5 reaches 72.9% on CursorBench; OpenAI proposes 'Useful Intelligence per Dollar' scorecard; Anthropic publishes agentic AI risk assessment framework; 4 frontier models launched in 8 days.

1. 1. Sora 2 Achieves Indistinguishable Deep Video Cloning

   Sora 2's deep video cloning captures every facial muscle movement. A single frame is indistinguishable from real footage.

2. 2. Kimi K3 Tops Frontend Coding Leaderboard, Open-Source Challenger

   Kimi K3 scores 1,679 on Frontend Code Arena, beating Claude Fable 5 and GPT-5.6 Sol. The 2.8T-parameter open-source model wins 6 of 7 sub-categories.

3. 3. NVIDIA Releases Nemotron 3 Embed Series

   NVIDIA releases Nemotron 3 Embed with three open-source checkpoints. 8B version ranks #1 on RTEB with 78.46 average NDCG@10.

4. 4. Tongyi's Wan-Streamer v0.2: 550ms End-to-End Multimodal Model

   Tongyi Lab releases Wan-Streamer v0.2, unifying listening, seeing, speaking, and acting in a single Transformer with 550ms response latency.

5. 5. Apple Sues OpenAI, Legal Battle Escalates

   Apple sues OpenAI alleging multiple misconducts, sends legal letters to ~40 former employees. Analysts link the move to Apple's new Siri AI competition.

6. 6. Moonshot AI Reveals Kimi K2.5 Technical路線 at GTC 2026

   Moonshot AI CEO reveals Kimi K2.5 at GTC 2026: MuonClip optimizer (2x data efficiency), Kimi Linear attention (beats full attention at 1M tokens), and Agent Swarm architecture.

7. 7. Schema Harness Reaches 99% on ARC-AGI-3 Public Set

   Schema Harness scores 99% RHAE on ARC-AGI-3 public set with Claude Opus 4.8 and Fable 5, and 95.35% with GPT-5.6 Sol, without modifying model weights.

8. 8. Claude Fable 5 Hits 72.9% on CursorBench

   Cursor's eval lead confirms Claude Fable 5 reaches 72.9% on CursorBench (Max effort mode), excelling on SWE-bench and other benchmarks.

9. 9. OpenAI Proposes 'Useful Intelligence per Dollar' Scorecard

   OpenAI proposes 'Useful Intelligence per Dollar' metric, measuring AI ROI across useful work completed, actual cost, and result reliability.

10. 10. Anthropic Publishes Agentic AI Risk Assessment Framework

    Anthropic's deputy CISO shares agentic AI risk framework with four core questions evaluating data access scope, tool permissions, and result impact.

11. 11. 4 Frontier Models in 8 Days: AI Race Heats Up

    Grok 4.5, GPT-5.6, Muse Spark 1.1, and Kimi K3 all launched within 8 days, intensifying competition on the Artificial Analysis Intelligence Index.