Eight OpenAI voice models run until January — and the price of four of them can no longer be looked up

On 20 July 2026 OpenAI notified developers that its legacy audio, realtime and transcription models leave the API on 20 January 2027. The list names eight model families: gpt-realtime, gpt-audio, gpt-realtime-mini, gpt-audio-mini, gpt-4o-realtime, gpt-4o-audio, gpt-4o-mini-realtime and gpt-4o-mini-audio, plus one transcription snapshot. Those models are still in service and still billed. Their published prices, however, are gone. We checked each of them today against OpenAI's own developer site. The four models in the gpt-4o branch no longer have a documentation page at all: platform.openai.com/docs/models/gpt-4o-realtime, gpt-4o-audio, gpt-4o-mini-realtime and gpt-4o-mini-audio all return 404. There is no context window, no modality list and no rate card to consult for a model a developer may be paying for tomorrow morning. The four newer models keep their pages, but not their full price. The API price list at platform.openai.com/docs/pricing now shows only two entries in the realtime and audio section — gpt-realtime-2.1 at $32 and $64 per million audio tokens, and gpt-realtime-2.1-mini at $10 and $20. None of the eight retiring models appears there. Their own model cards carry a pricing table headed "Text tokens" and nothing else: GPT-Realtime mini is listed at $0.60 input, $0.06 cached and $2.40 output per million tokens, GPT-Audio mini at $0.60 and $2.40. For models whose entire purpose is speech, the audio rate — the one that dominates any real bill — is not stated anywhere. The schedule has been tightening for a while. The original October 2025 snapshots of both mini models were shut down on 23 July 2026, three days after the deprecation notice, leaving only the December 2025 snapshots running. And the minis are themselves replacements: they became the recommended path on 7 May 2026, when gpt-4o-mini-realtime-preview and gpt-4o-mini-audio-preview were removed. Eight months later they are on the list too. None of this breaches any commitment — OpenAI publishes shutdown dates and names a successor for every entry, which is more than most of the industry does. But there is a difference between announcing an end date and quietly removing the ability to check what a running model costs. Anyone still calling these endpoints has five months left and no rate card to plan against. wujec.ai now carries profiles of all four surviving-until-January models in the current family, each with its shutdown date recorded.