GPT-Live-1.1 Replaces ChatGPT's Voice Model Today. The API Still Sells Only GPT-Live-1.
OpenAI switched ChatGPT Voice to GPT-Live-1.1 for Plus and above (including Codex and dots) and GPT-Live-1.1 Quip for Free and Go on October 15, 2026, with 13 new voices across French, German, Hindi, Japanese, Korean, Mexican Spanish and English. There is no blog post, no capability benchmark, and no API: the only Live model developers can buy is still gpt-live-1 at $0.05 per minute, and the same-day native Decisions integration in the API is tagged to gpt-live-1, not 1.1. We read the release notes, the forum post, the system card, the Voice help article, the developer docs, and Artificial Analysis.
By T.W. Ghost
OpenAI changed the model behind ChatGPT Voice today, October 15, 2026. GPT-Live-1.1 is now the default voice model on Plus and higher plans, including voice in Codex and voice calls with dots, and GPT-Live-1.1 Quip, which the release note calls "a mini version of the model", is the default for everyone else. They replace GPT-Live-1 and GPT-Live-1 mini, the pair that launched on July 8. The deliverable is 13 new voices across French, German, Hindi, Japanese, Korean, Mexican Spanish and English, plus what OpenAI calls "more natural accents across languages."
There is no openai.com blog post. The announcement lives in a ChatGPT release-note entry, a system card, the ChatGPT Voice help article, and a post on OpenAI's developer forum that embeds a video from OpenAI's own YouTube channel. That paper trail contains no accuracy, latency, or accent measurement of any kind.
The catch is bigger than the paperwork. This is a consumer rollout only. The forum post says "Rolling out today to all users. Coming soon to the API." No model ID, no date, no price. The only Live model a developer can buy as of tonight is still gpt-live-1 at $0.05 per minute. And the API changelog entry dated today that reads like a 1.1 feature, "native Decisions integration in GPT-Live", is tagged to gpt-live-1 and gpt-6-luna. Two releases, not one. All of it lands a day after Grok Voice 3.0 took the top of Artificial Analysis' Speech to Speech Index, where OpenAI's best row is still the old model. We read the release notes, the forum post, the twelve-section system card, the help article, the developer docs, and the AA leaderboard. Here is what held up.
Two Models, Thirteen Voices, No Numbers
| GPT-Live-1.1 and GPT-Live-1.1 Quip | |
|---|---|
| Announced | Oct 15, 2026 (ChatGPT release notes, system card, forum post); rolling out to all users |
| Replaces | GPT-Live-1 (Plus and above, Codex, dots) and GPT-Live-1 mini (Free, Go) |
| Surfaces | ChatGPT Voice in Live mode, Codex voice in the desktop app, voice calls with dots |
| API model ID | - (none; "Coming soon to the API", no date given) |
| API price | - for 1.1 and Quip; gpt-live-1 stays at $0.05 per minute, billed per second, backend tokens extra |
| Context / max output | - (not in the system card; no docs page exists) |
| Knowledge cutoff | - (not in the system card; the gpt-live-1 model page says Jul 31, 2025) |
| Accuracy, WER, latency, accent metrics | - (OpenAI published none) |
| Modalities | Audio and text; Live does not support video or screen sharing |
| Architecture | Full-duplex two-model system; search and reasoning are delegated to a backend text model |
| New voices | 13: two each for six non-English languages, plus Bubbie in English |
| Independent benchmarks | - (Artificial Analysis lists GPT-Live-1 only) |
The voices are the one concrete deliverable. The help article's "Additional voices for Live" table names Aube and Rivage (French), Sonna and Miko (German), Leher and Neer (Hindi), Himari and Rindō (Japanese), Haetsal and Hanul (Korean), Tinta and Arrullo (Mexican Spanish), and Bubbie (English). The count of 13 comes from the forum post; the help article labels no voice as new, and its table also lists Viola and Rio for Brazilian Portuguese, which today's release note does not mention, so we treat those two as pre-existing. The nine remastered English voices from July (Arbor, Breeze, Cove, Ember, Juniper, Maple, Sol, Spruce, Vale) remain. Two caveats from OpenAI's own pages: "The voices you see may vary by region and account", and "Changing the selected voice during a Voice conversation starts a new voice call in the same chat."
The accent claim is a sentence, not a measurement. The forum post promises "considerably better performance in French, German, Hindi, Mexican Spanish, Korean, Japanese, and many other non-English languages." The release note says "more natural accents across languages" and that Live "uses a wider range of short phrases to let you know when it's searching", which the forum post frames as less "checking" filler. No accent rating, filler rate, word error rate, or latency figure exists in any primary source. Dashes for all four.
Delegation and steerability are a forum sentence. "Delegation is now more natural, and steerability has also improved" appears in the forum post and nowhere else: not in the system card, the release note, or the help article. No metric. Unverified.
The backend is only partly disclosed. The card describes GPT-Live as a two-model system with a separate "thinking" model. Since September 9, per the release notes, "ChatGPT Voice can now use GPT-5.6 or GPT-6 Astra when it needs to search or reason through harder questions", picked with the same model and effort controls as text chat (see our GPT-6 coverage). OpenAI has not restated that for 1.1; the card's alignment evaluations use GPT-5.6 Terra as the underlying model, and the API docs suggest gpt-6-luna or gpt-6-sol as backends.
The Plan Table Is the Only Price List
The help article, updated today, is the closest thing to a price sheet. It puts Free and Go on Quip and names GPT-Live-1.1 on the Plus, Pro, and Business Standard rows; the system card's rule puts every higher tier on 1.1 as well. In Chat mode, Live usage "is measured over a rolling 24-hour period."
| Plan | Included Live time |
|---|---|
| Free (Quip) | - ("Limited access", no hours stated) |
| Go (Quip) | 3 hours |
| Plus | 3 hours |
| Pro ($100/month) | 15 hours |
| Pro ($200/month) | Unlimited, "subject to abuse guardrails" |
| Pro 500 ($500/month) | - (no Voice row in the help article or in the Sep 29 note that created the plan) |
| Business Standard / Premium | 3 / 15 hours, then 1.25 credits per minute |
| K-12 Teachers, Legacy Enterprise and Edu | 3 hours |
| Enterprise, Edu, Clinicians (credit-based) | 1.25 credits per minute |
| Enterprise (usage-based) | $0.05 per minute |
The hours are unchanged from the September 9 note that set them for GPT-Live-1, and that note also ended the fallback: "Plus and Pro no longer switch to GPT-Live mini after reaching a Voice limit." Two numbers deserve a second look. Enterprise usage-based billing is $0.05 per minute, the same figure as the API rate for the old model, and the ChatGPT Learn pricing page says "Voice in Desktop uses your existing Codex usage budget at $0.05 per minute", with the task model billed separately at token rates. A call can also end for reasons with no published number: "A Voice conversation may end when you reach a usage limit, the maximum session length, or when a long conversation reaches its context limit." Maximum session length and context limit are both dashes.
The Benchmarks, Sorted by Who Reported Them
House rules: independent numbers first, stated as fact; OpenAI's numbers labeled as OpenAI's.
Independent: nobody has measured it. Artificial Analysis' Speech to Speech Index (which carries no version label; the text-model Intelligence Index is still v4.3.2 and does not cover voice) has no row for GPT-Live-1.1, Quip, or GPT-Live-1 mini. It has GPT-Live-1 under two backends and, as of yesterday, a new leader.
| Model (Artificial Analysis, Speech to Speech Index) | Index | tau-Voice agentic | Cost per hour of input audio |
|---|---|---|---|
| Grok Voice 3.0 (xAI, now SpaceXSI; AA lists it as SpaceXAI) | 86.9 | 71.6% | $4.80 |
| GPT-Live-1 (Astra, medium) | 83.2 | 74.5% | $5.83 |
| Gemini 3.8 Live Extended Thinking (High) | 82.6 | 68.6% | $3.50 |
| Grok Voice 2.0 (deprecated by xAI) | 81.3 | 56.5% | $4.80 |
| GPT-Live-1 (Sol, low) | 80.1 | 59.3% | $4.47 |
Grok Voice 3.0 leads on the index, speech reasoning (99%), and task success (94.8%), with a time to first audio of 0.81 seconds against 1.34 seconds for GPT-Live-1 (Astra, medium). The one column OpenAI still owns is the agentic one: 74.5% on tau-Voice against Grok's 71.6%. Three AA caveats travel with these numbers. Both GPT-Live-1 rows are "based on 1 trial" on the tau-Voice component. The Arena Elo (1056 and 1057 for the two OpenAI rows, 1070 for Grok) "is frozen when a model first becomes eligible for publication". And AA's cost per hour counts audio and text in both directions plus "separately exposed reasoning/thinking tokens", which is why $5.83 and $4.47 sit above the $3.00 per hour (our arithmetic) that $0.05 per minute implies for the session alone. Whether 1.1 moves any of this is unknown, because OpenAI published nothing to compare.
Vendor: the only table OpenAI published is a safety table. The system card contains safety and alignment evaluations and no capability numbers. What follows are OpenAI's numbers, higher is safer, each cell printed as the predecessor's score to the successor's score. OpenAI's reading of the synthetic table, in full: "We observe that the GPT-Live-1.1 models perform close to parity with their predecessors. Analysis shows that GPT-Live-1.1 has statistically significant regressions on sexual, mental health, emotional reliance, and gore categories compared to GPT-Live-1; more research is needed to understand the regressions, and we plan to improve in these categories in future releases. These evaluations do not account for system-level interventions that elevate the holistic safety of the assistant response, such as Trusted Contact, localized crisis helplines, and Parental Controls for younger users."
| Category (synthetic prompts, generated audio) | GPT-Live-1 to 1.1 | GPT-Live-1 mini to Quip |
|---|---|---|
| Sexual | 0.87 to 0.84 | 0.83 to 0.83 |
| Illicit behavior | 0.94 to 0.93 | 0.92 to 0.92 |
| Mental health | 0.70 to 0.63 | 0.60 to 0.59 |
| Personal data | 0.95 to 0.95 | 0.93 to 0.94 |
| Emotional reliance | 0.86 to 0.81 | 0.80 to 0.80 |
| Self-harm | 0.89 to 0.87 | 0.85 to 0.86 |
| Hate | 0.96 to 0.96 | 0.93 to 0.94 |
| Gore | 0.76 to 0.69 | 0.66 to 0.62 |
On production prompts, real user audio shared with permission and de-identified, the card says: "We observe that the GPT-Live-1.1 models perform at parity with their predecessors, with no statistically significant regressions."
| Category (production prompts, real user audio) | GPT-Live-1 to 1.1 | GPT-Live-1 mini to Quip |
|---|---|---|
| Sexual | 0.77 to 0.81 | 0.69 to 0.78 |
| Illicit behavior | 0.87 to 0.86 | 0.84 to 0.86 |
| Mental health | 0.78 to 0.81 | 0.76 to 0.81 |
| Personal data | 0.92 to 0.92 | 0.89 to 0.92 |
| Emotional reliance | 0.76 to 0.76 | 0.70 to 0.79 |
| Self-harm | 0.85 to 0.86 | 0.84 to 0.88 |
Three honest observations the forum post does not lead with:
- The two tables point in different directions. On synthetic prompts GPT-Live-1.1 is lower than GPT-Live-1 on six of eight categories and flat on two; the largest drop is mental health, 0.70 to 0.63. On production prompts it is equal or higher on five of six, lower only on illicit behavior by 0.01. Quip is higher than GPT-Live-1 mini on every production row and lower on synthetic gore (0.66 to 0.62) and mental health. Which table matters to you depends on whether your risk is the typical call or the unusual one, and the card attaches its system-level-safeguards qualifier to the synthetic one.
- Do not line these up against July. The GPT-Live-1 column in today's card (synthetic sexual 0.87, mental health 0.70, gore 0.76) does not match the corrected GPT-Live-1 values in the July card (0.95, 0.85, 0.93). Same model name, different numbers, no explanation from OpenAI. The July card was itself corrected on August 4 because "the results had been generated using a backend configuration that did not match the final model release", a rerun that disclosed six statistically significant regressions against the originally published figures. Within-card comparisons only.
- The alignment results are pictures. Respecting Warnings, Blocker Deception, and Broken Search Tool are published as figures with no values in the text. The qualitative verdicts: on Respecting Warnings "the GPT-Live-1.1 models have slight regressions compared to their predecessors"; on Blocker Deception "GPT-Live-1.1 performs slightly better and GPT-Live-1.1 Quip performs slightly worse than their respective predecessors"; on Broken Search Tool the reverse. The comparison baseline is GPT-5.6 Terra. On the Preparedness Framework, "neither, when operating without delegation, could plausibly be considered High in any of our Preparedness Framework's Tracked Categories", with bio, chem, and cyber safeguards inherited from whichever backend model the work is delegated to.
Two Releases, Not One: The Decisions Hook Belongs to gpt-live-1
The API changelog entry dated today reads in full: "Added native Decisions integration in GPT-Live. Your app can act while your user is speaking or react to what Live says. Let Live call Decisions directly through native delegation, or trigger decisions as input or output transcript deltas arrive." Its tag line is "Feature · Model: gpt-live-1 · Model: gpt-6-luna · API: v1/live/sessions". The "Connect voice to Decisions" docs page says "Native delegation skips a round trip through your app", and its code sample sets "model": "gpt-live-1" with "decisions": {"model": "gpt-6-luna"}. The 1.1 system card never mentions Decisions.
So the feature is real, and it belongs to the model you can already buy. The Decisions API itself has been in public beta since October 6: gpt-6-luna is "the only model currently available", input costs $0.10 per 1M tokens with no cache or output charges, OpenAI claims typed answers "about 10x faster than the Responses API", and GA is expected "in the coming weeks." Two things no page states: whether a native delegation call from a Live session is billed at that Decisions rate on top of the $0.05 per minute session fee, and any millisecond figure for the latency it saves. Our own seed brief merged the voice-model update and the Decisions hook into one launch. The primary sources do not, and neither should you.
The Only API Price Is for the Old Model
The gpt-live-1 model page says "Voice sessions cost $0.05 per minute, billed per second", with backend model and tool usage charged separately; concurrent-session limits are 50 on Build, 300 on Launch, and 500 on Grow. Regional data-residency and FedRAMP endpoints add a 10% uplift for models released on or after March 5, 2026, which works out to $0.055 per minute (our arithmetic; OpenAI prints no regional figure). The model went GA in the API on September 10 and carries no deprecation notice.
The "cheaper mini tier" some readers will infer from Quip does not exist for developers. GPT-Live-1 mini was never an API model, and Quip is not one either; the mini story is a ChatGPT-plan fact for Free and Go. The cheapest OpenAI voice endpoint remains gpt-realtime-2.1-mini from the Realtime family, billed per token: audio $10.00 in / $0.30 cached / $20.00 out and text $0.60 / $0.06 / $2.40 per 1M, against $32.00 / $0.40 / $64.00 audio for gpt-realtime-2.1. One naming trap: gpt-live-transcribe at $0.017 per minute is a GPT-Live-branded speech-to-text model on the transcription endpoint, not a voice model.
Where You Can Run It
ChatGPT Voice in Live mode on web, iOS, Android, and the desktop app; voice in Codex inside the desktop app, where GPT-Live has handled Chat, Work, and Codex since July 23; and voice calls with dots, which roll out to Pro, Business Premium, and Enterprise (an off-by-default beta) and exclude the EEA, Switzerland, and the UK for Pro at launch. ChatGPT Voice is available in eligible Business, Enterprise, Edu, and Healthcare workspaces, a change from July's "not available at launch" that OpenAI never dated in the release notes. "Live does not support video or screen sharing"; the Advanced mode ("the previous real-time Voice experience") and the turn-by-turn Standard mode remain when Live is unavailable by plan or settings. Plugins work in Live since September 23. Audio clips are stored with the transcript and retained for 30 days. Not in the API.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Rolled out Oct 15, 2026 in ChatGPT Voice, Codex, and dots; Plus and above on 1.1, Free and Go on Quip | Verified (ChatGPT release notes, system card, Voice help article) |
| 13 new voices, two each for six languages plus Bubbie | Verified (count from the forum post; names from the help article, which labels no voice as new) |
| "Considerably better" non-English performance and less "checking" filler | Unverified (OpenAI's wording; no accent, WER, filler, or latency metric is published anywhere) |
| Delegation more natural, steerability improved | Unverified (forum post only; absent from the card, release note, and help article; no metric) |
| GPT-Live-1.1 or Quip in the API | Unverified ("Coming soon"; no model ID, date, or price; no docs page for either) |
| Native Decisions hook is a GPT-Live-1.1 feature | Contradicted (API changelog tags the entry Model: gpt-live-1, Model: gpt-6-luna; docs sample uses gpt-live-1; the 1.1 card never mentions Decisions) |
| A cheaper mini voice tier now exists | Contradicted for the API (GPT-Live-1 mini was never an API model and Quip is not one; true only for ChatGPT Free and Go) |
| Four statistically significant synthetic-prompt regressions vs GPT-Live-1 | Verified (system card Table 2, printed with OpenAI's "close to parity" lead-in and system-level qualifier; production prompts show no significant regressions) |
| GPT-Live-1 scores 83.2 (Astra, medium) and 80.1 (Sol, low) on AA's Speech to Speech Index | Verified (Artificial Analysis; both rows single-trial on tau-Voice, Arena Elo frozen at first publication) |
Who Should Use It
Use GPT-Live-1.1 if you already pay for Plus or above and talk to ChatGPT in French, German, Hindi, Japanese, Korean, or Mexican Spanish, where the new voice pairs land, or if you drive Codex by voice in the desktop app. You do not get a choice in any case: it is the default, and the production-prompt safety table gives no reason to avoid it. Budget the same three hours per rolling day as before.
Skip it for building anything. There is no API, no model ID, no price, and no date; a product plan written against "coming soon" is a plan written against a forum reply. Buy gpt-live-1 at $0.05 per minute if you need OpenAI's full-duplex stack now, and read the native Decisions integration as a gpt-live-1 feature. Skip Live altogether if your workflow needs video or screen sharing, which stays in Advanced mode, and treat any "X% better accents" claim you see this week as unsourced, because OpenAI published none.
The frontier is elsewhere: on Artificial Analysis' Speech to Speech Index, Grok Voice 3.0 leads at 86.9 as of yesterday, with GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live Extended Thinking (High) at 82.6 for $3.50 per hour of input audio, the cheapest of the three. OpenAI still holds the agentic column at 74.5% on tau-Voice. Whether 1.1 closes the gap or widens it is a question only AA can answer, once OpenAI ships something AA can test. Our model comparison has the current standings side by side.
Update, October 16, 2026
- Oct 16: Early user reports, not vendor-confirmed: a forum moderator split five posts from the announcement thread into a new topic reporting wrong voices, interruptions, and language-recognition errors in Standard and Advanced voice modes since the 1.1 rollout. The first report is timestamped the evening of October 15; by midday October 16 the topic had 11 posts and no staff reply. OpenAI's release notes do not acknowledge it, and the launch-day status incident (ChatGPT mobile logins and signups, 04:29 PM to 06:56 PM) does not list Voice, Live, or the API.
- Oct 16: A forum moderator posted the "Connect voice to Decisions" explainer as additional material, again describing the integration for "GPT-Live" with the
gpt-live-1sample, which is the reading we print above. - Oct 16: The developer side did move. OpenAI's API changelog released prompt-to-voice for all API customers: create a custom voice from a text description through the API or the voice Playground, creating the voice is free, and speech runs on regular GPT-Live session billing. The same entry added nine natural voices for GPT-Live: Nuri, Juni, and Flitz for German; Brise and Sillage for French; Nira for Hindi; Harema and Shida for Japanese; Noeul for Korean. The docs still use
gpt-live-1style model IDs rather than agpt-live-1.1path, the Decisions API was still beta that day, and Artificial Analysis had added no GPT-Live-1.1 or Quip row. - Oct 16: Press coverage in The Neuron and Notebookcheck matched the primary sources on date, tiers, voices, and the safety split; neither printed a price, latency, or benchmark figure, and neither should be cited for one. Headlines saying "six languages" are counting languages, not the 13 voices.
Sources
- OpenAI: ChatGPT release notes (October 15, 2026 entry)
- OpenAI Developer Community: GPT-Live-1.1 is rolling out with new voices and better accents
- OpenAI: GPT-Live-1.1 system card
- OpenAI: GPT-Live-1.1 system card, voice-native evaluations on synthetic prompts
- OpenAI: ChatGPT Voice help article (plan table and voice names)
- OpenAI developer docs: gpt-live-1 model page
- OpenAI developer docs: API changelog
- OpenAI developer docs: Connect voice to Decisions
- Artificial Analysis: Speech to Speech Index
*Not sure whether a voice model you cannot buy yet should change your plans, or whether the one you can buy is the better bet? Take the free 2-minute quiz and get matched. Then read about the model that took the top of the voice leaderboard the day before: Grok Voice 3.0.*