What actually retires on 11 December
Two modes go at once. Standard Voice is the old turn-taking system: you speak, it waits, it answers. Advanced Voice is the natural-sounding mode OpenAI introduced in 2024, the one that could handle interruptions and laughter and the occasional sigh. Both are still selectable in Settings today.
After the cutoff, GPT-Live is the only path. OpenAI sent formal notices to users and published release updates on 10 November, which is where the 31-day window comes from. The company says voice usage has nearly doubled since GPT-Live debuted in July, and that virtually all active voice traffic has already migrated. That migration rate is the reason the legacy modes are being switched off rather than kept as an option.
Advanced Voice reached the ChatGPT desktop app on macOS and Windows in October 2024, as Dataconomy reported at the time. The retirement itself is set out in OpenAI's own help-centre notice, “Retiring Advanced and Standard Voice in ChatGPT.”
The video and screen-sharing exception
Advanced Voice does not disappear entirely on day one. It stays alive for two features only: live video and screen sharing, because GPT-Live does not handle those yet. OpenAI's notice gives no date for when it will.
So there is a bridge, but it is narrow. It is not a way to keep talking to the old conversational mode. It is a way to keep pointing your phone camera at things while the assistant describes them.
The hourly allowances in one table
| Plan | Voice allowance |
|---|---|
| Go and Plus | 3 hours per 24 hours |
| Pro, $100/month (≈ €86) | 15 hours |
| Pro, $200/month (≈ €172) | Unlimited |
Euro figures use an exchange rate of about $1.16 to the euro; OpenAI bills in dollars, so EU customers also carry card and conversion fees that vary by bank.
OpenAI turned these into explicit hourly buckets in September, replacing vaguer daily limits.
Run the arithmetic on the middle tier: 15 hours against €86 works out to roughly €5.75 per hour of talk time if you use all of it. Voice is priced as a subscription, not as a utility.
Transcripts survive, audio does not
Two retention clocks are running here, and they are different. Past voice conversations remain in chat history as text, and the retirement of a mode does not touch them. Stored audio and video clips tied to those conversations are deleted after 30 days.
If you use voice to draft notes, dictate a report or capture a meeting summary, the transcript is the durable artifact. The recording is not. Anyone who relied on the audio as the source of truth should pull what matters before the 30-day window closes on it, independently of what happens on 11 December.
What a transcript costs to process elsewhere
Once the audio is gone, the transcript can go anywhere. That is the more interesting number for developers, because voice subscriptions are expensive and tokens are not.
Take one hour of English speech. At roughly 150 words per minute that is about 9,000 words, which lands near 12,000 tokens at the usual 1.3 tokens per word. Summarise it into a 600-token brief and the bill looks like this:
| Model | Input | Output | Per hour of talk |
|---|---|---|---|
| DeepSeek-V4.1-Flash (off-peak) | $0.15 / 1M | $0.60 / 1M | ≈ $0.0022 |
| Mistral Large 4 | $0.70 / 1M | $2.10 / 1M | ≈ $0.0097 |
| GPT-6.1 Terra | $1.00 / 1M | $5.00 / 1M | ≈ $0.015 |
Those are my own calculations from published per-token prices and a stated assumption of 12,000 input and 600 output tokens. Change the summary length and the totals move, but not the conclusion: summarising one hour of speech costs between about a fifth of a cent and one and a half cents, depending on the model. Compared with that one hour of summarised speech, the €86 to €172 monthly subscription is roughly 5,000 to 80,000 times larger — about three to five orders of magnitude. The exact ratio depends on the pairing: €86 ÷ $0.015 is about 5,700×, while €172 ÷ $0.0022 is about 78,000×.
The other option is to keep it local. In AI Arena we track tokens per second and cost per token side by side on an RTX 5060 Ti, and for batch transcript summarisation the local path is not about saving fractions of a cent. It is about not shipping the transcript to a third party at all.
Disclosure rules and the 30-day clock
The regulatory side changed this year. Under Article 50 of the EU AI Act, transparency obligations for AI systems that interact with humans are now legally binding rather than voluntary. A company that puts a ChatGPT voice agent in front of European customers has to make clear that the customer is talking to a machine, and that has to be in the interaction itself.
A mode migration is a practical moment to check that the disclosure is still in the flow and not buried in a footer that nobody reads. Nothing in the 11 December change removes the requirement.
Retention is the other item for a record of processing. A human voice is personal data, and the 30-day clip deletion is the figure to document. OpenAI's retirement notice does not say where voice data is processed or stored, which is a question for your data processing agreement rather than for a release note.
Will my old voice conversations disappear on 11 December?
No. Transcripts stay in chat history. Audio and video clips attached to them are deleted after 30 days, which runs on a separate clock from the mode retirement.
Can I still choose Advanced Voice after the cutoff?
Only for live video and screen sharing, and only until GPT-Live gains those features. OpenAI has not published a date for that handover.
Does GPT-Live cost extra?
There is no separate charge. It draws on the voice allowance of your plan: 3 hours per 24 hours on Go and Plus, 15 hours on the $100 Pro tier, and unlimited on the $200 tier.
The one item without a date attached is the video and screen-sharing handover. Advanced Voice stays switched on for those two features until GPT-Live catches up, and OpenAI's notice stops there.
