Mobile dictation with OpenAI transcription through the environment's Codex account

[ignaciomontenegroc]

Problem

iPhone dictation runs on Apple's on-device SpeechTranscriber. It's convenient and private, but for coding prompts it falls short of OpenAI's transcription models: library names, file paths, CLI flags and mixed-language speech often come out wrong, so I end up fixing the draft by hand. Non-English speech has its own problems (#12743, #13313), and phones without iOS 26 or Android have no dictation at all (#13603).

Proposal

Add OpenAI transcription as an opt-in engine for mobile dictation. It runs on the environment the phone is already connected to and uses the Codex/ChatGPT account the environment is already signed into. The phone never needs an API key, which keeps to the product decision in #15422 and the design note on #8718 (credentials live on the environment, not the phone).

• Setting: Voice input > Transcribe with: This iPhone (default, unchanged) or Environment (OpenAI via Codex).
• Flow: the phone records as it does today, uploads the finished recording to the environment, and the transcript goes into the draft for review. Nothing is sent automatically, and the audio is deleted afterward, as it is now.
• Availability: the option only appears when the environment has a signed-in Codex provider that supports transcription.
• Fallback: if the environment is unreachable, use on-device transcription, or show a clear error.

How this relates to existing ideas

• Mobile dictation through the user's own environment server #20725 proposes the same environment path with a local Parakeet model. OpenAI transcription could be a second engine on that pipeline instead of separate work.
• Would Codex-backed live dictation for web and desktop be in scope? #12212 proposes Codex-backed live dictation for web and desktop, routed through the thread's environment and Codex account. This asks for the mobile version, reusing that account routing.

Would an environment-side OpenAI engine, using the existing Codex sign-in, be in scope for mobile dictation?