One app, separate experiences
OpenAI's help documentation makes the distinction explicit: the ChatGPT desktop app brings Chat, Work and Codex into one application, but Codex keeps its separate view, workflows and history. The July 23 update lets eligible accounts on macOS and Windows start tasks by voice in either Work or Codex, ask about progress and direct work within the tools and permissions of the selected experience.[1], [2]
The live transcript is useful for review, but the documentation says only one Voice conversation can run at a time. That is a clear limit for teams that want to run more than one Voice conversation in parallel. The meaningful test for me is accepted-output rate, correction time and permission exceptions on a fixed task set; the launch documents report neither those outcomes nor a verifiable Codex-only weekly-user count.[1]
Routing is valuable only when it can explain itself
Runway's Media Router matches an image, video or audio request against cost, latency and quality preferences, with price caps and provider allow and deny lists. Through one endpoint it returns the selected engine and metadata explaining why it was chosen. If no option satisfies the constraints, it returns an explicit error instead of silently degrading them; a dry-run mode shows the choice without generating anything.[3]
Voice and routing add two layers above the model: one turns intent into a task, the other turns a media request into a generation engine. The developer bottleneck may therefore move from producing an output to supervising the choice. A fixed prompt suite with blinded quality, end-to-end latency, total cost and failure rate would reveal which layer actually saves work. For now, the clearest option value lies in seeing why a choice was made and retaining the ability to switch providers.[3], [2]