OpenAI’s coding tool Codex just got a real upgrade: Subagents. In this video I break down what that actually means in practice—Codex can now spin up multiple specialized agents in parallel, each handling a different part of the job (security, code quality, bugs, race conditions, test flakiness, maintainability), and then merge everything into one coherent output. The key point is you trigger it in plain language, and Codex handles the coordination for you in the background. I also go through the custom agent system where you define agents via config files—model choice, reasoning effort, sandbox mode, tool/MCP access, and instructions—so you can stop wasting your strongest model on “lookup work” and reserve it for the heavy lifting. Then Sam Altman posts Codex growth charts and takes a pretty direct shot at Claude Code, framing Codex as what “serious builders” are switching to. The chart is basically a hockey stick from early 2026 onward, and the timing lines up with features like subagents that push Codex beyond autocomplete-style workflows. That said, I’m not convinced it’s as one-sided as the tweet implies—Claude Code is a real competitor, the models feel different, and the workflows matter (Codex is more of its own environment, while Claude Code sits closer to the terminal-native flow many devs already prefer). Finally, Google drops practical Gemini API cost controls in AI Studio: project-level monthly spend caps (with a small delay caveat), automated tier upgrades, and new dashboards for rate limits and cost breakdowns. Not flashy, but exactly the kind of tooling you want if you’ve ever been surprised by an API bill or rate-limited at the worst time.