Maya: Hey builders! Here's what's on the radar today on Codex Insider. Codex Sites gets scheduled refreshes and per-person data, OpenAI adds a twelve-hour prompt cache, Sites lands custom domains and private tunnels, Codex CLI ships a searchable config, and Higgsfield opens an API. I'm Maya. James: And I'm James. Let's get into it. Maya: First up, Codex Sites, where you build sites by asking Codex, got two features. You can now link a site to a scheduled task that refreshes its data automatically, instead of creating those tasks separately and losing track of them. James: Right, and the second one is the piece I keep thinking about, Maya. It's called bring your own plugin, and a site pulls from each visitor's own app connectors, so two people open the same site and see different data. Maya: Exactly right, and OpenAI's example is a team mission control site. It lists shared priorities, then shows each teammate the pull requests they need to review or their assigned Linear tickets, after they grant access. To try both, OpenAI says just ask Codex. James: Story two. OpenAI added an optional twelve-hour prompt cache to the API, the way programs send requests to its models. Shared instructions, tools, and context stay cached across requests hours apart, and reused content is billed at the cheaper cached-input rate. Maya: That's the key framing, James, it's aimed at persistent agents that resend the same instructions every call. For logged-in users, Codex and Work reportedly go further and keep the cache with no time limit, so resuming an old conversation carries no penalty. James: And separate figures for Codex say cached input is now permanent, with reads ninety to ninety-five percent cheaper hit more often, and cache writes never charged. OpenAI's announcement does not state those figures, and the API cache is optional, so no existing request gets it by default. Maya: Story three, more for Sites. On October fifteenth OpenAI highlighted two organization features in posts on X. Workspace custom domains let you publish a Site on a web address your organization has verified it owns, like support dot sites dot yourdomain dot com. James: And the second, private tunnels, is the part I'd flag, Maya. Hosted Sites can connect to your company's approved internal infrastructure and the APIs those systems offer, meaning a Site reaches your own systems once they've been approved. Maya: Exactly, and this came with a wider Enterprise update. Sites is now generally available for Enterprise, with US and EU data residency for new Sites in eligible workspaces. OpenAI describes it as building team apps with databases and authentication, and linked its Sites guide for beginners. James: Story four. OpenAI released Codex CLI zero point one six three point zero on October fifteenth. Three changes touch daily use. First, slash config is now searchable, so you look up preferences across categories instead of hunting, and the screen flags which changes need a restart. Maya: And the one I love, James, is steering mid-turn. If Codex is heading the wrong way, you send a follow-up while it's still responding and the notes say it redirects immediately, no waiting the turn out. Interrupting is separate, and it stops active Code Mode cells too. James: That's the standout, Maya. Third, Codex can now run independent read-only tools in parallel, including skill, memory, and conversation-history lookups, though the notes give no timing figures, so the speedup isn't quantified. Also worth checking: voice mute moved to F nine, and the leader key defaults to Control X. Maya: Last story. Higgsfield, the AI image and video platform, opened an API, a doorway letting other software send it work. Tools like Claude Code and Codex can now send image and video requests straight to Higgsfield and get the finished files back, no tab-switching. James: And in practice, Maya, one test fed old YouTube thumbnails into Codex, asked it to study colors, expressions, and spacing, then it produced a new thumbnail following those patterns. The same brief in Claude Code with Opus five point five came back with different framing to compare. Maya: Setup is a one-time step per project: create an API key, have the agent write your environment file with placeholders, then paste credentials and save. One thing to watch, James, the API is billed separately, you add funds and pay for what you use, video priced per second. James: Good caveat, Maya, and that pay-as-you-go part means you skip a plan like Max, which started at one hundred thirty-nine dollars a month. New accounts saw fifteen percent off at recording, and Higgsfield says it does not train on API data. Offers can change, so confirm what your account shows. Maya: That's your Codex Insider daily brief. Check codexinsider.ai for the full stories. Keep building.