Three places it shows up
1. The bottom of a fallback chain
The text brain of my AI receptionist, written in April 2026, tried Llama on Groq first, then Gemini 2.5 Flash, then gpt-4o-mini. If all three failed, the caller got an apology. OpenAI was last because it was the one I paid full price for, and the job did not need more than a small model.
In the months that code ran, the order mattered less than having three. Groq retired the first model in August. A chain with one link would have gone silent.
2. The first voice on my phone line
Version one of the phone receptionist, dated April 2, 2026, used OpenAI's tts-1 for speech. Twilio gathered the caller's words, Groq wrote the reply, OpenAI turned the whole reply into an MP3, and Twilio played the file. The voice was fine. The call was slow, because the caller waited for a complete audio file on every turn. I pre-recorded five filler phrases like "One moment" and "Let me check on that" to cover the gap.
I replaced it with a streaming pipeline: Twilio Media Streams, Deepgram for listening and Cartesia for speech, sending audio in 80-millisecond chunks as it was generated. The swap had little to do with voice quality. Streaming into a phone call needs 8 kHz mu-law audio over a WebSocket, and Cartesia handed me that format directly. The Cartesia page has the details.
3. The request format
This is the part of OpenAI I use every week. Groq, OpenRouter, DeepSeek and Together all accept the same /chat/completions request OpenAI defined. In an ebook tool I built in November 2026, one adapter file covers all of them. Switching providers is a base URL and a model name. The adapter also carries a price table and refuses to run without a daily dollar cap set, a habit I picked up after a billing outage.
OpenAI's open-weight models reach my production systems the same way. openai/gpt-oss-120b, hosted by Groq at $0.15 in and $0.60 out per million tokens, is a live fallback in two of my products today. So an OpenAI model does answer my customers. OpenAI just does not bill me for it.
Prices on November 14, 2026
| Model | Input per 1M | Cached input | Output per 1M |
|---|---|---|---|
| gpt-6-astra | $10.00 | $1.00 | $50.00 |
| gpt-6.1-sol | $2.00 | $0.10 | $10.00 |
| gpt-6.1-terra | $1.00 | $0.05 | $5.00 |
| gpt-6-luna | $0.10 | $0.01 | $0.50 |
| gpt-5-nano | $0.05 | $0.005 | $0.40 |
| gpt-4o-mini | $0.15 | $0.075 | $0.60 |
Speech is priced apart from text. tts-1 costs $15 per million characters and tts-1-hd costs $30. Transcription with gpt-transcribe is $0.0045 a minute, and Whisper is $0.006 a minute. For comparison, Groq hosts Whisper large-v3-turbo at $0.04 an hour, which is under a tenth of a cent per minute.
What I can't tell you
I have not run OpenAI's realtime voice models, its agent tools or fine-tuning in a product. If you need a judgment on those, ask someone who has.