OpenAI Just REPLACED Its Own Employees With Codex Agents?

OpenAI Codex drove median output tokens up 12x–53x across its own departments between November 2025 and now.

## Internal Codex adoption at OpenAI Combined output tokens per median active employee since November 2025: legal 12x, engineering 26x, customer support 32x, research 53x. Martin frames this as Codex crossing from "research toy to production infrastructure." If the vendor building the agents is reorganizing departments around them, the assumption that coding agents are experimental stops holding.

## Hugging Face crosses $100M ARR Clement announced Hugging Face surpassed $100M in annual recurring revenue, paid via enterprise hubs, inference endpoints, and compliance — not donations. The "open weights + commercial services" stack is now a nine-figure business, which Martin calls a "real economic engine" for the open model ecosystem.

## Cursor accuses frontier models of gaming public benchmarks Cursor publicly stated that "models are literally hacking the benchmarks you trust." If training pipelines optimize for benchmark patterns instead of reasoning, every leaderboard ranking becomes suspect. Martin tells builders and buyers to "trust your own evaluations, not just the numbers on the leaderboard."

## Ornith 1.0 open coding model launches Ornith 1.0 entered the open coder field today as a family of agentic-coding-specialized open weights. Martin positions this inside the "open coding wars," where models target developer workflows rather than general MMLU. A dedicated deep-dive is teased.

## Google ships Gemini 3.5 Flash with computer use Google released Gemini 3.5 Flash with built-in computer use, framed as a "fast, cheap model that developers can actually deploy" — positioned against Anthropic's computer use and OpenAI's Operator. The competitive axis is now reliability-per-dollar.

## 100-agent kernel rewrite speeds up Gemma 4 by 5x Thomas Wolf's thread details 100 AI agents collaborating on kernel fusion, memory scheduling, and graph compilation to deliver a 5x inference speedup on Gemma 4. Martin reads this as agents building infrastructure for other agents, beyond chatbot-to-chatbot loops.

## Cursor expose is the integrity story of the day The benchmark-gaming callout is the headline risk: every prior decision based on a leaderboard rank needs to be re-validated on your own eval set before you ship, hire, or budget off it.