GPT-Live-1.1 Rolled Out on October 15: Six New Languages, Two New Models, and One Uncomfortable Document

The least interesting thing about GPT-Live-1.1 is the thing OpenAI led with. The X post on October 15, 2026 announced "new voices, better accents, and less 'checking...'" — three claims about how a voice model sounds. The genuinely interesting thing is in the companion artefact almost nobody reads on launch day: the GPT-Live-1.1 system card, published the same day, which names a successor model for the mini tier (GPT-Live-1.1 Quip), reassigns which model different ChatGPT tiers get by default, and reports small but statistically significant regressions against GPT-Live-1 on a specific class of evaluations.

That is a lot of structure hidden behind a post about accents.

What shipped, and to whom

The rollout is a consumer-surface change first. GPT-Live-1.1 and GPT-Live-1.1 Quip are rolling out in ChatGPT Voice to all users as of October 15, 2026. The release notes and the voice help page agree on the shape of it: thirteen new voice options, two each for French, German, Hindi, Japanese, Korean and Mexican Spanish, plus Bubbie in English. Live also picks from a wider range of short phrases to signal that it is searching, which is the fix behind the announcement's "less checking…". The help page adds the caveat that matters for support: the voices you see may vary by region and account, and you choose them under Settings → Voice, then Voice.

The API is explicitly not part of this. The announcement says "coming soon to the API," and the documentation backs that up: the models page lists gpt-live-1 and gpt-live-transcribe and nothing else in the Live family, and gpt-live-1.1 is a 404. There is no model ID, no price, no rate limit. Anyone who tells you what GPT-Live-1.1 costs per minute is quoting GPT-Live-1's $0.05-per-voice-minute figure and hoping you do not check.

The part that matters for anyone building: the tier split

The system card lays out the architecture plainly. GPT-Live-1.1 is the successor to GPT-Live-1. GPT-Live-1.1 Quip is the successor to GPT-Live-1 mini. And the default assignment has changed: GPT-Live-1.1 is the default voice model for Plus and higher tier users — including for Codex and Dots — while GPT-Live-1.1 Quip is the default for everyone else.

Read that again, because it means the product most people will experience is not the model in the headline. Free-tier and lower-tier users are getting Quip. If you are evaluating "GPT-Live-1.1" from a Free account, you are evaluating a different model than a Plus user is, and the system card treats them as separate evaluated systems throughout.

The card also restates the structural fact that gives this whole family its shape: GPT-Live is a two-model system, and the authors hypothesise that GPT-Live models "would inherit their alignment behaviour from their 'thinking' models" — the text models they delegate to. That is not a footnote. It is the load-bearing assumption behind the safety evaluation, and it is why the card compares results against the underlying model used for those evaluations, GPT-5.6 Terra.

Where the eval evidence is solid, and where it is not

The headline safety result is genuinely reassuring, and it is narrow. On voice-native evaluations with production prompts — the real traffic distribution — GPT-Live-1.1 and GPT-Live-1.1 Quip "perform at parity with their predecessors, with no statistically significant regressions." That is the number that speaks to what users will actually encounter.

On synthetic-prompt evaluations, the picture differs, and the card says so directly: both models show statistically significant regressions on sexual, mental health, emotional reliance, and gore categories compared to GPT-Live-1. There is also a slight regression on Respecting Warnings. On Blocker Deception, GPT-Live-1.1 performs slightly better and Quip slightly worse. On a broken-search-tool scenario, GPT-Live-1.1 performs slightly worse and Quip slightly better.

The honest reading is not "the models got less safe." It is that the two evaluation modes disagree, and the card tells you why the authors still consider the production-prompt result the operative one: synthetic prompts probe corners that ordinary voice traffic does not reach. But this is exactly the kind of detail that does not survive the trip into launch coverage, and it is worth holding onto — a version bump inside a model family is not automatically a safety-neutral event, and here the direction of movement on adversarial probes was negative.

On capability limits, the card is unusually concrete. Neither model, operating without delegation, could plausibly be considered High in any of OpenAI's Preparedness Framework Tracked Categories — and the reason given is architectural rather than a matter of restraint: these models lack broad access to tools independently of the models they delegate to, and have no code execution capability. The safety argument and the capability ceiling are the same fact stated twice.

What the six languages actually buy

OpenAI's framing is that accents in non-English languages are more natural. There is no public, reproducible benchmark behind that claim — no index row, no vendor figure with a method attached, just the assertion and, now, a set of voices you can listen to yourself.

Where independent measurement exists, it exists for the predecessor. On the Artificial Analysis Speech to Speech Index, GPT-Live-1 appears twice — once at 83.2 with the Astra backend at medium reasoning effort, and once at 80.1 with the Sol backend at low effort — against Gemini 3.8 Live Extended Thinking at 82.6 and Grok Voice 3.0 at 86.9. There is no GPT-Live-1.1 row on that board. The index is a composite of speech reasoning, conversational dynamics, agentic performance, arena preference and task success, and it is revised as models are added and re-run; the numbers above are from a capture taken the day after launch.

So for now the accent claim is a vendor claim, and the only independent evidence available is for the model it replaces. That is not a criticism — it is simply the state of the information on day one, and it will change as the board catches up.

• What is verified — the rollout, the thirteen new voice options, the tier defaults, the system-card evaluation results, the absence of an API model ID.

• What is vendor-reported only — better accents, more natural delegation, and the wider range of search-signalling phrases.

• What does not exist yet — any GPT-Live-1.1 row on an independent leaderboard, any per-minute price, any third-party latency or cost-per-hour measurement.

The layer underneath, and who is exposed to it

Everything above describes a voice front-end. The reasoning in these sessions runs on a text model, and both the system card and the API documentation are explicit that the voice layer and the reasoning layer are separately metered and separately substitutable. The card names GPT-5.6 Terra as the underlying model used for its evaluations. That is the layer where cost variance lives, where quality differences between configurations show up, and where switching is cheap because it does not touch the voice integration.

OrcaRouter carries more than 200 models behind a single key at provider list price with no markup, so an upstream price change lands on our side the same day rather than at the next contract renewal. To be plainly accurate about what this article is and is not: we do not host GPT-Live-1.1, GPT-Live-1.1 Quip or GPT-Live-1 — those voice endpoints come from OpenAI's own API, and the 1.1 versions are not in any API at all yet. The text models these systems hand work to are a different story, and that is where automatic failover and the routing DSL do something real: you can point delegation at a chosen backend, swap it on live traffic, and keep a call alive when one provider errors.

What to watch

Two dates matter more than the launch. The first is whenever gpt-live-1.1 gets a model ID and a price — that is the moment this stops being a consumer story and starts being a procurement decision, and the number will be compared against GPT-Live-1's $0.05 per voice minute from the moment it appears. The second is the next revision of the Speech to Speech Index, because that is when the accent and delegation claims either acquire independent support or do not.

Until then, the useful posture is the boring one: if you are on Plus or above you already have the new model and can judge the accents yourself; if you are on a lower tier, you have Quip, and the difference between the two is the single most under-reported fact of this release. And if you are building, nothing about your stack changed on October 15, 2026 — which is itself information.

A result like this is also a reminder that model mistakes are already documented: the AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.