🤖 AI HOT Daily · Aug 15, 2026

Xiaohongshu open-sources dots3-note Preview, a 280B lightweight model for long-horizon agents and multimodal reasoning; Zhipu releases GLM-5.3, open-source No.1 in coding with emerging cyber-security skills; DeepSeek V4 Pro lands on SiliconFlow with a 1M context; Cursor is officially acquired by SpaceX, gaining the world's largest GPU cluster; Claude will add text watermarking to comply with the EU AI Act; OpenAI and Anthropic wage a price war against Chinese rivals; OpenRouter data shows 84% of tokens come from non-frontier models.

1. 1. dots3-note Preview Open-Sourced

   The lightest dots3 model — 280B total, 16B active — supports 512K context and text/vision/audio understanding, tuned for complex reasoning and long-horizon agent tasks.

2. 2. GLM-5.3: No.1 Open-Source Coding, Emerging Cyber Skills

   Built on GLM-5.2's base with extreme post-training scaling, coding jumps 50%, topping open models on Terminal Bench 3.0 near Claude Fable 5, and matching Mythos 5 on white-box review.

3. 3. Gemini 3.7 Flash Rolls Out to Pro and Ultra

   Gemini 3.7 Flash is now available to Pro and Ultra users in chat, improving multi-step reasoning and accuracy, with Gemini Spark also on 3.7 Flash.

4. 4. DeepSeek V4 Pro Arrives on SiliconFlow

   Day-0 support with a 1M context window and low/high/max reasoning intensities, emphasizing coding, tool calling, and agent workflows — still MIT-licensed.

5. 5. Cursor Officially Acquired by SpaceX

   Completing the acquisition that began in April, Cursor gains the world's largest GPU cluster to build stronger models for less; Grok 4.6 is an early fruit of the tie-up.

6. 6. How Claude's Text Watermarking Works

   Future Claude text will carry a watermark signaling it was AI-written, an EU AI Act compliance change built on SynthID-Text with no real impact on output quality.

7. 7. OpenAI and Anthropic in a Price War

   The price war makes API cost elasticity real; users who fled to Chinese vendors over bills may now re-compare capability and price after the cuts.

8. 8. Indonesia's First University AI Center Opens

   The ministry, Indosat, NVIDIA, and Gadjah Mada University launched the UGM Indosat NVIDIA AI Technology Center to nurture local AI talent.

9. 9. Summer 2026 Open-Model Ecosystem Review

   HF public model repos grew to 2.96M, yet 85.6% have under 200 downloads; Chinese labs' largest monthly open models span 754B–2.78T params, far ahead of US labs.

10. 10. Maximizing Token Value in Claude Code Sessions

    Running /clear between tasks trims stale context; switching models or effort mid-session busts the prompt cache and drives up cost.

11. 11. Who Needs SOTA? 84% of Tokens Come from Non-Frontier Models

    OpenRouter data shows the six most-used models perform at ~77% of frontier levels for just 2.5% of Claude Fable 5's cost; top open models keep closing the gap.