Friday, August 14, 2026

The Brief

The era of using AI just to draft polite emails is closing. OpenAI just demonstrated ChatGPT Work, showing how the model can aggregate context from scattered documents to generate fully structured internal project hubs. It is a clear pivot from text generation to infrastructure management. This aligns with new OpenAI research confirming that leading enterprises are ditching basic chatbots for autonomous agentic systems. The gap is widening between companies using AI to brainstorm and those using it to execute.

Raw speed is the catalyst for this shift. OpenAI previewed an Ultrafast tier for GPT-5.6 Sol running at 14x speed. Powered by Cerebras hardware, the model spits out 750 tokens per second. Analytical tasks that used to drag on for hours now take minutes. Google is sprinting to match this pace, launching Gemini 3.7 Flash a mere three weeks after its 3.6 update. But the rapid release cycle might not be enough. New market data indicates Gemini is losing significant usage share to both ChatGPT and Claude. Trust and model performance are outweighing Google's built-in ecosystem advantage.

Speaking of Claude, Anthropic wants its assistant in every corner of your workflow. The Claude in Chrome side panel has been rebuilt as a Cowork session that syncs your conversation history and skills across web, desktop, and mobile. You can start a complex data pull in a browser tab and finish it on your phone. Anthropic is also leaning hard into autonomy. Starting today, Claude Code defaults to Auto mode for premium users. As we noted earlier this week, Anthropic is betting that human intervention is often just a bottleneck in modern development.

But passing the keys to the machines introduces serious risks. Attackers just exfiltrated terabytes of credentials in a massive supply chain attack targeting users of a compromised AI software package. If you let an agent write and execute code, you better be certain about what it imports. Adding to the security headache, researchers figured out how to reverse-engineer original LLM prompts from output text with near-perfect accuracy, without needing access to the model weights. If your company relies on hidden system prompts to protect its intellectual property, consider those secrets spilled.

While everyone else races to build new agents, Microsoft is finally cleaning house. The company is merging its consumer and business Copilot apps and stripping out underperforming features like Deep Research and AI-generated podcasts. Sometimes the smartest product update is just deleting the noise.

Bottom line: Your role is shifting from manual execution to high-level orchestration. As tools like ChatGPT Work and Claude Code take over building infrastructure and writing code, your true value will lie in quality control and security auditing. Lock down your API keys, audit your dependencies, and assume your clever system prompts will eventually be read by everyone.