(aiBytes #3) Kimi-K3, cheaper frontier models & agents that break in

⚡ aiBytes_ #3

Good morning, builders 👋

Welcome back to your Monday-morning dose of aiBytes_. I had this issue half-written as a price-cut roundup - three labs, three cheaper models, tidy week. Then Anthropic published that its own models had breached three companies during security tests, and Sam Altman, of all people, said he's ready to decelerate. Back to the drawing board 😅

Capability is outrunning containment: Kimi-K3, DeepSeek-V4-Flash and GPT-5.6 all made frontier intelligence cheaper, while Anthropic's position on open-weights models set off the loudest argument of the week over who should be allowed to hold it.

Today: Moonshot's open-weights flagship and the argument it reopened, two price cuts in 48 hours, DeepMind putting whole-body control in a robot, and seven repos worth your weekend.

🔥 Kimi-K3 on HuggingFace - Moonshot's new open-weights flagship took 1,378 points and 545 comments, the top AI thread of the week. Story of the week (Hugging Face)

📜 Our position on open-weights models - Anthropic laid out where it stands, and 1,746 comments piled in behind it. (Anthropic)

⚡ DeepSeek-V4-Flash Update - another price-performance jump, shipped straight into the API changelog. (DeepSeek)

💸 Advancing the price-performance frontier with GPT‑5.6 - OpenAI's answer landed the day after, same argument, different curve. (OpenAI)

🦾 Gemini Robotics 2 brings whole body intelligence to robots - DeepMind moves past manipulation to whole-body control. (DeepMind)

🔐 Anthropic says its own AI models breached three companies during security tests - after OpenAI's models broke into Hugging Face, Anthropic checked its history and found three more. (TechCrunch)

🐌 Sam Altman is ready to decelerate - the reversal follows "the first security incident that I have felt very viscerally." (TechCrunch)