The convenience of a single text-to-audio endpoint can conceal a lifecycle risk. OpenAI’s speech documentation as reviewed on October 25, 2026 explicitly lists older speech models for retirement on January 6, 2027, so deployment architecture now matters more than a one-off voice sample.
What the older endpoint does
The documented POST /v1/audio/speech endpoint converts supplied text to an audio output; the guide illustrates gpt-4o-mini-tts with voice selection and natural-language delivery instructions. It also describes tts-1 and tts-1-hd as alternatives with different latency/quality design goals. Output formats and voices vary by model. Do not assume every advertised voice works with each identifier, or that a speech endpoint can perform conversation state, tool calls or interruption handling.
The dated lifecycle caveat
On the checked documentation, tts-1, tts-1-hd and gpt-4o-mini-tts are deprecated and scheduled to stop on January 6, 2027. The page directs migration to a Realtime model, explicitly noting that this is a change of API workflow rather than a drop-in model-string replacement at /v1/audio/speech. Archive the checked notice with build documentation and validate any later vendor change; do not present this future shutdown as already completed in October 2026.
Use the right synthesis test
For prerecorded narration, use the same neutral script and pronunciation checklist on each eligible voice. Test 1,000-character passages as well as short prompts, confirm format and playback compatibility and measure the actual audio duration. For an interactive experience, benchmark the connection setup, stream startup and user interruption path, then test loss of connectivity. A streaming audio response is not automatically an interactive speech-to-speech agent.
Disclosure is part of the API use case
OpenAI’s guide says its usage policies require clear disclosure to end users that the voice is AI-generated rather than human. Include the disclosure in the player, interface or surrounding programme material where a reasonable listener will encounter it. Protect any speech reference files, avoid imitation of people without a valid authorized pathway and separate listening consent from logging consent.
When to choose another stack
Use a specialized narration product if voice-directed acting and long-form studio editing outweigh integrated model workflows. For a browser agent with handoffs and interruptions, assess the Realtime API from the beginning. A hybrid system can keep an editorial master in one studio and produce short status prompts using a separate real-time system, provided the voice identity and licensing permit that use.
Worked production example
A website that already generates narrated article audio with /v1/audio/speech should create a migration inventory immediately. List every application call site, scheduled job, retry worker, cache key and voice used by the old integration. Test a Realtime alternative on isolated sample content, not customer traffic, because connection lifetime, event framing, response cancellation and audio format may differ. Keep a reversible release switch if the new product is not a faithful replacement for a static MP3 generation workflow. The January 2027 date is a future scheduled event, not an observed outage.
Practical acceptance gate
A migration acceptance test needs at least five outputs per supported language: numbers, names, long punctuation, rapid snippets and a full-length article. Capture total generation cost under the new billing unit and audio-processing budget, but publish monetary figures only after the pricing team verifies the tariff. Check that listeners receive the AI-origin disclosure and that expired caches can be rebuilt. Revalidate this warning at the actual deployment date in case the vendor changes or extends the notice.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: availability varies by account/endpoint and jurisdiction. Version: gpt-4o-mini-tts / tts-1 / tts-1-hd. Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Open Tts — official documentation (checked 2026-10-25)
- Open Api — official documentation (checked 2026-10-25)
- Open Rt — official documentation (checked 2026-10-25)