Matthew Berman opens with OpenAI's preview of GPT-5.6 running on Cerebras hardware. A coding demonstration drops from more than twelve minutes to under two, shifting the practical bottleneck from model reasoning to tool calls, local CPUs and the human cost of supervising many parallel agents.
The video then examines Anthropic's planned text watermarking. Claude would bias low-stakes token choices using a secret key so generated passages can be detected later, while lightly edited human text and most functional code offer less room for the signal. Matthew Berman questions whether even small changes to token selection can remain completely neutral for users.
Grok 4.6 is presented as a cheaper model moving close to the frontier, and Grokbot turns that capability into a simpler agent interface with plugins, delegation and shared conversations between agents. The broader pattern is that capable agent systems are becoming easier to operate without exposing every model and tool decision.
GLM 5.3, DeepSeek V4 Pro and Meta's smaller Muse Glimmer model illustrate the improving open-model field. The larger releases approach leading coding benchmarks at much lower token prices, while Muse Glimmer targets on-device use rather than frontier performance. The video closes with OpenAI's opt-in computer-history feature, which can observe selected applications and suggest automations but will require careful privacy choices. Subscription and cross-promotion segments are omitted.
Watch on YouTube
