GPT-Live-1.1 API pricing — $0.05 a minute, $0.02 on standby
GPT-Live-1.1 API pricing is 0.05 dollars per minute of active session time and 0.02 dollars per minute on standby. OpenAI made the voice model available to developers on November 19, 2026 (US time) under the model ID gpt-live-1.1 at the v1/live/sessions endpoint. Voice sessions are billed by duration, not by token. The standby rate applies only to WebRTC sessions while inference is paused, and silence alone does not trigger it. The backend model that handles reasoning and tool calls behind the voice is billed separately at its own rates. Structured outputs, fine-tuning and image input are not supported. The voice catalog, the list of supported languages, latency figures and concurrency limits for 1.1 have not been published.
The three lines
- Price — $0.05 per active minute ($3 an hour) and $0.02 per standby minute ($1.20 an hour), per the OpenAI model page. Billing is by session duration; per-second billing is stated by two secondary sources
- Conditions — standby applies to WebRTC sessions while inference is paused. Without an idle timeout the session stays at the active rate. Backend model and tool usage is billed separately
- Limits — no structured outputs, fine-tuning or image input. Voices, languages, latency and 1.1 concurrency limits are unpublished. Sources give the API release date as November 19 or November 20, 2026
Key questions
- GPT-Live-1.1 API pricing
- **$0.05 per minute of active session time and $0.02 per minute on standby. Backend model and tool usage is billed separately.** | Item | Rate | Condition | |---|---|---| | Active session | $0.05 per minute | Session open with inference running | | Standby | $0.02 per minute | WebRTC sessions only, while inference is paused | | Backend model and tools | That model's rates | Separate from the voice session charge | | Per hour | $3.00 active / $1.20 standby | This article's arithmetic |
- GPT-Live-1.1 standby pricing how it works
- **The session must use WebRTC and inference must be paused. A quiet stretch on its own does not switch the session to the standby rate.** | Condition | Detail | Source | |---|---|---| | Transport | WebRTC | OpenAI model page | | State | Inference paused (session.paused) | Model page, Data Studios | | Setting | session.idle_timeout_s as a whole number from 5 to 60 seconds | Data Studios, DEV Community post | | If omitted | Stays active after the first wake | Data Studios | | WebSocket and SIP | Not addressed on the model page; the DEV post says active rate only | Unsettled |
- GPT-Live-1.1 vs GPT-Realtime
- **GPT-Live-1.1 is a full-duplex model that listens while it speaks and bills by the minute. GPT-Realtime takes turns and bills by the token.** | Item | GPT-Live-1.1 | gpt-realtime-2.1 | |---|---|---| | Conversation | Simultaneous listening and speaking | Turns detected by voice activity | | Billing | $0.05 per minute ($0.02 standby) | Audio at $32 in / $64 out per 1M tokens | | Architecture | Voice front end that delegates to a backend model | One model hears, reasons and speaks | | Standby rate | Yes (WebRTC) | No | | Image input | Not supported | Listed as supported in the DEV post | The GPT-Realtime figures come from a single DEV Community post.
GPT-Live-1.1 costs $0.05 per minute of active session time in the API, and $0.02 per minute on standby. OpenAI opened the voice model to developers on November 19, 2026 (US time). The model ID is gpt-live-1.1 and the endpoint is v1/live/sessions. The OpenAI model page describes it as a voice model that listens while it speaks, and says voice sessions are billed by duration while backend model and tool usage is billed separately.
Two things will be unfamiliar to developers used to text models. The unit is the minute, not the token. And there are two rates: while a user pauses to think, or a shop kiosk waits for the next customer, a session can drop to a lower standby price. That lower price has conditions. This piece sets out the price table, when standby actually applies, worked costs for different kinds of service, how the model differs from GPT-Realtime, and what OpenAI has not yet published.
1. The price table — $0.05 and $0.02 a minute, plus a backend bill
The table separates what the OpenAI model page states from what secondary sources add.
| Item | Detail | Basis |
|---|---|---|
| Model ID | gpt-live-1.1 (the only snapshot, and the default) | OpenAI model page |
| Endpoint | v1/live/sessions (Live API) | Model page, Data Studios |
| Active session rate | $0.05 per minute | Model page and three other sources agree |
| Standby rate | $0.02 per minute | Model page and three other sources agree |
| Standby condition | WebRTC sessions while inference is paused | Model page |
| Billing basis | Session duration | Model page |
| Per-second billing | Per second, no rounding up to the minute | Data Studios, DEV Community post |
| Session start | A WebRTC session bills 15 seconds up front, credited against later duration | Data Studios, DEV Community post |
| Backend model and tools | Billed separately at that model's normal rates | Model page and three other sources agree |
| Token, cache and batch discounts | Do not apply to the voice session charge | Data Studios |
| Modalities | Audio and text in, audio and text out | Model page |
Converted into more familiar units, by this article's arithmetic:
| Unit | Active | Standby | Difference |
|---|---|---|---|
| 1 second | about $0.00083 | about $0.00033 | — |
| 1 minute | $0.05 | $0.02 | 60% lower |
| 1 hour | $3.00 | $1.20 | $1.80 |
| 8 hours | $24.00 | $9.60 | $14.40 |
| 15 seconds at session start | $0.0125 | — | Credited against later use |
Data Studios, citing OpenAI cost documentation, gives a simple formula: billable voice seconds ÷ 60 × the per-minute rate, plus backend cost. The first half is settled by the tables above. The second half is the harder one to forecast.
GPT-Live-1.1 does not do everything itself. According to the DEV Community post, it is a voice front end that hands reasoning and tool calls to a text model behind it. There are two ways to do that: a responses mode in which OpenAI runs the backend, and a client mode in which the developer's application runs it. The mode cannot change mid-session. Two secondary sources name gpt-6-luna as a starting backend and gpt-6-sol for harder tasks. Those names are not on the model page.
So the invoice has two lines. One scales with how long the call lasted. The other scales with how much thinking the conversation required. A call that exchanges greetings and a call that queries a booking system three times can be the same length and still differ on the second line. In other words, the per-minute figure is a floor on the bill, not the bill.
2. The $0.02 standby rate — silence alone does not earn it
Standby is the headline number of this release, and the easiest to misread. Superpower Daily notes that a quiet stretch does not qualify by itself; the session has to be in the paused state.
| Condition or setting | Detail | Basis |
|---|---|---|
| Transport | WebRTC session | Model page |
| State | While inference is paused | Model page |
| Wake policy | session.wake_policy.type set to "speech" | Data Studios, DEV post |
| Idle timeout | session.idle_timeout_s, a whole number from 5 to 60 seconds | Data Studios, DEV post |
| Voice detection threshold | session.wake_policy.vad_threshold, 0 to 1, default 0.5 | Data Studios |
| If the timeout is omitted | Inference stays active after the first wake and never drops to standby | Data Studios |
| State events | session.paused and session.resumed | Data Studios |
| Connection during standby | Connection and session ID persist | Data Studios |
| Muting the microphone | Does not close the session, so billing continues | Data Studios |
| First response after standby | May be slower; no figure is given | Data Studios |
Four points need care.
First, there is no standby unless you configure it. A session with no idle timeout pays $0.05 for every one of the ten minutes a user spends away from the desk. Standby is a feature the developer switches on, not a default.
Second, the timeout itself is billed as active. With a 10-second timeout, the 10 seconds after speech stops are still active time. Six silences mean 60 seconds billed at the active rate. Since the minimum timeout is 5 seconds, every silence costs at least 5 seconds of active time.
Third, mute is not standby. A muted microphone does not close the session or lower the rate. Standby is a server-side session state; mute is a client-side input block.
Fourth, WebSocket and SIP are unsettled. The model page says only that standby applies to WebRTC sessions. The DEV Community post states flatly that WebSocket and SIP sessions bill every connected second at the active rate, which is a stronger claim than the primary source makes. A contact center connecting over the phone network through SIP should confirm with OpenAI documentation before building a budget around standby.
The setting names above match across two secondary sources. Data Studios says its own code snippet is illustrative, not an official sample, and for the same reason this article does not reproduce a full configuration.
3. What your service would cost — worked examples
The first three cases are Data Studios calculations; this article re-checked the arithmetic. All of them cover the voice session charge only. Backend cost is excluded.
| Case | Session profile | Without standby | With standby | Saving |
|---|---|---|---|---|
| 30-minute tutoring session (WebRTC) | 12 minutes of talk, six 3-minute pauses, 10-second timeout | $1.50 | $0.99 | 34% |
| Shop kiosk, 8-hour shift (WebRTC) | 480 minutes open: 55 active, 425 standby | $24.00 | $11.25 | about 53% |
| Kiosk, 22-day month | The same shift, 22 times | $528.00 | $247.50 | about 53% |
| 6-minute support call | Continuous conversation | $0.30 | $0.30 | 0% |
The tutoring arithmetic runs like this. Each of the six pauses leaves a 10-second timeout billed as active, so active time is 12 minutes plus 1, or 13, and standby is 17. Thirteen minutes at $0.05 is $0.65, 17 minutes at $0.02 is $0.34, and the total is $0.99. For the kiosk, 55 × $0.05 = $2.75 plus 425 × $0.02 = $8.50 gives $11.25.
What the cases show is that the value of standby depends entirely on the share of silence. A six-minute call with no gaps saves nothing. A kiosk that waits for most of the day saves more than half. Even the kiosk never reaches zero: an eight-hour shift with no customers at all still costs $9.60 in standby if the session stays open. That may be worth comparing with a design that closes the session when nobody is there and opens a new one on approach. The deciding variables are how slow the first response after standby is and how long a fresh session takes to start. Neither has a published figure.
One larger-scale illustration follows. This is this article's own arithmetic, again excluding backend cost.
| Assumption | Monthly volume | Voice charge |
|---|---|---|
| One 90-second call | 1 call | $0.075 |
| 6-minute calls | 1,000 calls (6,000 minutes) | $300 |
| 6-minute calls | 10,000 calls (60,000 minutes) | $3,000 |
| 6-minute calls, 2 minutes of each on standby | 10,000 calls | $2,400 |
The last row assumes four active minutes and two standby minutes per call: 4 × $0.05 + 2 × $0.02 = $0.24, times 10,000. Real calls are messier. Silence comes in many kinds, from a customer hunting for a document to a lookup running in a back-office system, and the timeout eats into every gap. The sensible first step is to measure the distribution of silence lengths in your own call recordings.
The $0.075 figure for a 90-second call comes from the DEV Community post. It holds if the secondary sources are right that billing is per second with no rounding up. That detail was not confirmed against OpenAI text.
4. How it differs from GPT-Realtime — minutes versus tokens
OpenAI now has two voice API tracks. The gpt-realtime-2.1 figures below rest on a single DEV Community post, which should be kept in mind throughout.
| Item | GPT-Live-1.1 | gpt-realtime-2.1 |
|---|---|---|
| Conversation | Listens while it speaks (full duplex) | Turns detected by voice activity |
| Barge-in | Handled without a setting | interrupt_response setting |
| Architecture | Voice front end delegates reasoning and tools to a backend | One model hears, reasons, calls tools and speaks |
| Billing unit | Minutes of session time | Tokens |
| Voice rate | $0.05 per minute active, $0.02 standby | Audio at $32 in, $64 out per 1M tokens |
| Text rate | Backend model's rates | $4 in, $24 out per 1M tokens |
| Cached input | Not applicable to the voice session | $0.40 per 1M tokens |
| Standby rate | Yes (WebRTC) | No |
| Context | 128,000 tokens per session, managed automatically | 128,000 tokens, 32,000 max output |
| Endpoint | v1/live/sessions | Realtime API |
Full duplex is a telecom term for a link on which both sides can talk and listen at once, like a phone call, as opposed to a walkie-talkie where one side speaks at a time. As the DEV post describes it, GPT-Live-1.1 processes incoming audio while it is generating output, so it can offer short acknowledgements and change course mid-sentence. Superpower Daily says the model decides whether to keep listening or answer, instead of following strict turn-taking. The Realtime line waits for voice activity detection to judge that the other party has finished.
A one-line cost comparison is not possible because the units differ. The DEV post works out that $0.05 buys about 1,560 audio input tokens on gpt-realtime-2.1, which is $0.05 divided by $32 per million and checks out. But none of the sources read for this article says how many audio tokens a minute of conversation produces. Without that, nobody can say which is cheaper in general. What can be compared is structure.
| Kind of service | Likely better fit | Why |
|---|---|---|
| Long silences, sessions held open for hours | GPT-Live-1.1 | It has a standby rate; Realtime does not |
| Costs that must be forecast from call time | GPT-Live-1.1 | The voice charge scales with time, though backend cost is a separate variable |
| Responses needed in a fixed schema | Check first | GPT-Live-1.1 does not support structured outputs |
| A single-model setup with nothing to wire together | Realtime | No backend delegation to configure |
| Natural acknowledgements and interruptions | GPT-Live-1.1 | Simultaneous listening and speaking |
This table is inferred from the published design, not measured. The only reliable test is to run the same call recordings through both APIs and compare the invoices.
5. What it supports and what it does not
| Status | Item | Basis |
|---|---|---|
| Supported | Streaming | Model page |
| Supported | Function calling | Model page; under responses delegation the tools go in delegation.responses.tools (DEV post) |
| Supported | Audio and text input and output | Model page |
| Supported | WebRTC, WebSocket, SIP inbound and outbound | DEV post; the model page does not list transports |
| Not supported | Structured outputs | Model page |
| Not supported | Fine-tuning | Model page |
| Not supported | Image input | Model page and three other sources |
| Not supported | Video input | Superpower Daily |
| Not supported | Predicted outputs | Data Studios |
| Unsupported endpoints | Chat Completions, Responses, Realtime, Batch, Assistants, fine-tuning, audio and image endpoints | Model page |
Limits that have been stated:
| Item | Value | Basis |
|---|---|---|
| Context | 128,000 tokens per session, managed automatically | DEV post |
| Instructions | Up to 16,384 tokens | DEV post |
| Maximum output | Not published | DEV post |
| SIP call length | Two hours per connected call | DEV post |
| Other session length limits | Not published | — |
| Concurrent sessions | GPT-Live-1.1 and GPT-Live-1 are counted separately, each across inference regions | Model page |
| Per-tier limits for 1.1 | Not published; check the organization limits page | Model page |
| For reference: GPT-Live-1 | 50 on Build, 300 on Launch, 500 on Grow, no Free tier | Relayed by secondary sources |
The lack of structured outputs matters for business automation. The voice model cannot be asked to return its answer in a fixed JSON schema. Values that need a format, such as a booking reference or an order, have to arrive as function call arguments or be handled by the backend model. Fine-tuning is also unavailable, so tone and business rules have to fit inside the instructions, capped at 16,384 tokens.
As improvements over GPT-Live-1, OpenAI cites more accurate transcription and a more consistent accent and voice across languages. Superpower Daily relays a post from the OpenAI developer account saying 1.1 completes tasks more reliably. None of these claims comes with a number, and no benchmarks have been published. The active rate is the same as the earlier model.
6. What has not been published — voices, languages, latency, limits
This is why the price table alone is not enough to make an adoption decision. Several of the values that determine the quality of a voice product are blank.
| Item | Status | Why it matters |
|---|---|---|
| Voice catalog | In no source | Whether a brand can pick a suitable voice |
| Supported languages | Described only as multilingual | No basis for judging any single language in advance |
| Latency | No figures | The main determinant of whether a conversation feels natural |
| First response after standby | Only that it may be slower | The price paid for the standby discount |
| Concurrency for 1.1 | Not published | Concurrent calls are the capacity of the service |
| Regional availability | No list | Where data is processed, and latency |
| Session length cap | Nothing beyond the two-hour SIP limit | Design of all-day kiosk sessions |
| Transcription gain | No figure | The case for migrating from GPT-Live-1 |
| Total cost including backend | Depends on model and usage | The actual invoice |
For teams building in a language other than English, the missing language list is the largest gap. Whether the model page's word, multilingual, covers a given language, and how well, can only be established by testing. At $0.05 a minute, ten 10-minute test calls cost $5 before backend charges, so the testing itself is cheap.
The safety information is single-sourced. According to Superpower Daily, OpenAI published a system card on October 15, 2026 and added material on the API voice models on November 19. With backend delegation disabled, the 1.1 API is said to score numerically higher than its predecessor on mental health, emotional reliance and self-harm tests, and at parity on sexual content, illicit behavior and personal data. OpenAI is reported to say the tests are deliberately hard and are not estimates of real-world rates. The system card link in that article is broken, so the card was not read for this piece.
7. Frequently asked questions
Is there no token charge at all? Not for the voice session, which is billed by time alone. But whatever the backend model does in reasoning and tool calls is billed at that model's token rates. Tokens do not disappear from the invoice; they move to a separate line.
Does the rate drop to $0.02 automatically when the user stops talking? No. The session must use WebRTC, an idle timeout must be set, and the session must actually enter the paused state. Without a timeout, the session stays at the active rate after the first wake.
Does muting the microphone stop the meter? No. Data Studios says muting does not close the session. To stop charges, end the session. To lower them, send it to standby.
Does standby apply to phone calls over SIP? That is not settled. The model page says only that standby applies to WebRTC. The DEV Community post says SIP and WebSocket pay the active rate throughout, but the primary source does not go that far. Phone-based services are safer budgeting as if standby were unavailable.
Are calls under a minute rounded up? Two secondary sources say billing is per second with no rounding. On that basis a 90-second call costs $0.075. The model page says only that sessions are billed by duration, so this needs confirming on the pricing page.
Is it cheaper than GPT-Realtime? It cannot be said in general. One bills by the minute and the other by the token, and the sources read for this article do not give the number of audio tokens in a minute of conversation. The most that can be said is that services with long silences are structurally better served by GPT-Live-1.1, which has a standby rate.
Should I migrate from GPT-Live-1? The active rate is the same. The improvements OpenAI cites, transcription accuracy and cross-language voice consistency, are not quantified. Because the two models have separate concurrency limits, both can run side by side during a migration. The statement that GPT-Live-1 also gets standby pricing from the same date comes from one source and needs checking.
Is this the same model as ChatGPT voice? Data Studios says GPT-Live-1.1 is now the default voice for ChatGPT Plus and higher tiers. That is one source, and it adds that the variant described for other users is not on the API pricing page.
Can I get responses as JSON? The voice model does not support structured outputs. Values that need a format should come through function call arguments or be handled by the backend model.
What does a dollar buy? Twenty active minutes or fifty standby minutes of voice session time, before backend charges. That is this article's arithmetic from the published per-minute rates.
8. What remains open and unconfirmed
| Item | Confirmed | Not confirmed |
|---|---|---|
| Rates | $0.05 active, $0.02 standby per minute (OpenAI model page) | The pricing page itself was not opened |
| Billing basis | Session duration (model page) | Per-second billing and the 15-second initial charge rest on two secondary sources |
| Standby condition | WebRTC, inference paused (model page) | Whether it extends to WebSocket or SIP; sources diverge |
| Standby settings | Names and ranges match across two secondary sources | Not checked against an official sample |
| Release date | The cited changelog entry and one report's publication date say November 19, 2026 | The DEV post says November 20; the model page has no date |
| Backend cost | That it is billed separately | Recommended models are named only by secondary sources; their rates were not checked |
| GPT-Realtime comparison | Rates and limits from the DEV post | The OpenAI pricing page; audio tokens per minute |
| Standby for GPT-Live-1 | One statement by Data Studios | Whether standby first appeared with 1.1 |
| Improvements | Transcription accuracy, cross-language consistency (model page) | Any figures or benchmarks |
| Safety evaluation | System card contents as relayed by Superpower Daily | The system card itself |
| Voices, languages, latency | — | None published |
| Concurrency | That 1.1 and 1 are counted separately | Per-tier numbers for 1.1 |
| ChatGPT rollout | One statement by Data Studios | An OpenAI announcement |
In short: the GPT-Live-1.1 API costs $0.05 per active minute and $0.02 per standby minute, and standby applies only when a developer sets an idle timeout on a WebRTC session and the session actually pauses. The more silence a service has, the more it saves; a call with no gaps saves nothing. The backend model is billed on top, so the per-minute figure is the floor of the bill. With voices, languages, latency and concurrency limits still unpublished, the remaining questions have to be answered with test calls.
For how token rates are built, see the August 15, 2026 explainer "What AI token pricing is — why output costs about 5× input"; for rates that change by time of day, the August 16, 2026 explainer "What peak pricing for AI APIs is — the same prompt, twice the cost"; and for the risks that come with synthetic voices, the October 7, 2026 explainer "What voice cloning is — how 3 seconds of audio copies a voice, and how to stop it".
Sources
- OpenAI developer docs — GPT-Live-1.1 model page
- Data Studios — OpenAI GPT-Live-1.1 API: $0.05/min active, new $0.02/min standby mode
- DEV Community — GPT-Live-1.1 vs GPT-Realtime: Which OpenAI Voice API to Use (2026-11-20)
- Superpower Daily — OpenAI brings GPT-Live-1.1 to developers with lower-priced standby sessions (2026-11-19)