GPT-Live-1.1 vs GPT-Live-1: The Version Bump Changed Who Gets Which Model

Nothing in the app told anyone that the voice model changed. That is the point of a version bump inside a family: it is meant to be invisible. GPT-Live-1.1 replaced GPT-Live-1 in ChatGPT Voice on October 15, 2026, and the only public documents that explain the swap are the system card published the same day and a release-notes entry in OpenAI's help centre. The X post announcing it mentioned voices, accents and less "checking…" — none of which tells you what a version number inside a model family is actually for.

Here is what it is for, in this case: the bump is a re-assignment of which model different accounts talk to, plus a successor for the small model, plus a set of evaluation results that moved in both directions. If you use the product, one of those changed for you this week. If you build on it, none of them did.

The two-model structure, and why the mini tier matters more than the headline

GPT-Live has always been a pair, not a model. GPT-Live-1 shipped as the flagship and GPT-Live-1 mini as the lighter sibling, and OpenAI's own consumer help documentation assigns them by subscription level. GPT-Live-1.1 keeps that shape exactly: GPT-Live-1.1 is the successor to GPT-Live-1, and GPT-Live-1.1 Quip is the successor to GPT-Live-1 mini.

What changed is the mapping.

• Plus and higher — GPT-Live-1.1, including inside Codex and Dots.

• Free and lower tiers — GPT-Live-1.1 Quip.

• Everyone else, in practice — the same voice list, the same Settings → Voice path, but not the same model underneath.

The consequence is worth stating flatly, because it is the single most under-reported fact of this release: two people comparing "GPT-Live-1.1" in the same week may not be comparing the same model. The system card evaluates GPT-Live-1.1 and GPT-Live-1.1 Quip as separate systems throughout, reports separate results for each against GPT-Live-1 and GPT-Live-1 mini respectively, and never presents a merged number. That is the correct way to do it and it means any claim of the form "GPT-Live-1.1 does X" needs a tier attached to be meaningful.

The split is also not new. Five weeks earlier, on September 9, 2026, OpenAI restructured GPT-Live usage limits by plan — three hours on GPT-Live-1 mini for Go, three hours on GPT-Live-1 for Plus, fifteen for the $100 Pro tier, unlimited for the $200 one — and deprecated the separate Instant/Medium/High Voice intelligence levels in favour of picking a backend model and reasoning effort directly. The version bump inherits that arrangement rather than creating it; what changed on October 15 is which model sits at each rung.

There is a second structural fact the card restates, which explains why the safety work looks the way it does. GPT-Live is a two-model system at the architecture level: the voice model handles the conversation and delegates hard queries to a text model. The card's authors hypothesise that GPT-Live models "would inherit their alignment behaviour from their 'thinking' models" — that the delegation target, not the voice layer, is where the safety-relevant reasoning happens. That hypothesis is load-bearing for everything downstream, and it is why the card benchmarks against GPT-5.6 Terra, the text model used for those evaluations.

What the evaluations say, in both directions

The headline result is reassuring and narrow. On voice-native evaluations using production prompts — the actual traffic distribution — both GPT-Live-1.1 and GPT-Live-1.1 Quip "perform at parity with their predecessors, with no statistically significant regressions." For anyone worried that a version bump quietly degraded the thing they use every day, that is the number that answers the question.

The synthetic-prompt evaluations say something different, and the card does not bury it. Both models show statistically significant regressions against their predecessors on sexual, mental health, emotional reliance and gore categories. 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.

Two readings of that are available and the card supports the first. Reading one: the production-prompt result is the operative one, synthetic prompts probe corners ordinary voice traffic does not reach, and the two evaluation modes disagreeing is a known artefact of adversarial-prompt construction. Reading two: the direction of movement on adversarial probes is negative for both models, and a version bump is not automatically safety-neutral.

Both are true. The reason it is worth keeping both in view is that the second one almost never survives the trip into launch coverage, and it is the one that would matter if a future version moved further in the same direction.

On capability limits the card is unusually concrete, and the reason is architectural rather than a matter of restraint. Neither model, operating without delegation, could plausibly be considered High in any of OpenAI's Preparedness Framework Tracked Categories — because these voice 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, and it is the clearest example in this release of GPT-Live's structure doing the work that guardrails would otherwise have to.

The numbers that do not exist yet, and the one that changed

GPT-Live-1 has a public footprint that GPT-Live-1.1 does not.

• There is no gpt-live-1.1 model ID on OpenAI's models page. The Live family still lists gpt-live-1 and gpt-live-transcribe only.

• There is no per-minute price. GPT-Live-1's $0.05 per voice minute is the only number in the family, and it belongs to the older model.

• There is no row on the Artificial Analysis Speech to Speech Index. GPT-Live-1 appears twice on the board captured October 16, 2026 — 83.2 with the Astra backend at medium reasoning effort and 80.1 with Sol at low effort — and the successor appears nowhere.

That last absence is the one to be careful about. It is tempting to read GPT-Live-1.1's quality as GPT-Live-1's 83.2, and the two models are the same family with the same architecture and a stated parity result on production prompts. It is still not a measurement of GPT-Live-1.1. What the card actually licenses is the claim that the successor did not regress on production prompts, which is a different and weaker statement than a score.

Why a version bump is a routing problem

Step back from the model comparison and the interesting thing about this release is that it is a deployment event wearing a model's name. OpenAI rewrote which tier gets which model, introduced a successor for the small tier, and left the interface identical. Everything that changed is an assignment.

That is exactly the shape of problem a router exists to solve, and it is worth being precise about where the boundary sits. The consumer swap is OpenAI's to make and you have no say in it. The delegated reasoning layer is a different matter: the card names GPT-5.6 Terra as the backend used for its evaluations, and the API documentation treats the voice front-end and the reasoning backend as separately configured and separately metered. That is the layer where a version change on someone else's schedule does not have to become a change on yours.

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 renewal. The pattern this release argues for is a pinned delegation target with an explicit fallback: the routing DSL lets you name the backend a session hands work to, swap it on live traffic without touching the voice integration, and keep a call alive when a provider errors. To be accurate about what we are and are 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. The text models these systems delegate to generally are on the router.

There is a second reason the delegation layer is where a version bump should be absorbed. The system card's own safety argument depends on the backend: GPT-Live inherits its alignment behaviour from the model it hands work to. If that is the mechanism, then the backend is not an implementation detail — it is the part of the system that determines what the whole thing does when a query gets hard. Owning that choice is worth more than owning the voice wrapper.

What to do with this

If you are a user, check your tier. Plus and above has GPT-Live-1.1; below that has GPT-Live-1.1 Quip. The voice list is the same either way, so the tier is the only thing that tells you which model is answering, and it is not surfaced in the app.

If you are on the API, nothing changed on October 15, 2026. gpt-live-1 is still the model ID, $0.05 per voice minute is still the rate, and the documentation is unchanged. Do not build a budget around a model ID that does not exist yet.

If you are planning for the API release, the three things worth watching are the model ID, the price, and the first independent index row. The system card has already told you the shape of the safety story and the tier structure; what it has not told you, and cannot, is how the thing sounds to a caller. That is the same gap that existed with GPT-Live-1 in July 2026 and closed only when an outside lab ran the evaluation.

And the version number itself is worth calibrating expectations against. GPT-Live-1 went from a consumer feature to a metered API over two months, and the successor is on the same path with the same two documents — a card and a release note — standing in for a launch post. That is what a mature model family looks like from the outside: the news stops being the model and starts being the assignment.

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.