OpenAI priced GPT-Live-1 at $0.05 a minute, the number Vapi charges to host a voice agent with no model in it. The meter runs through silence, hold music and every second the backend is thinking, and the backend is a second bill.
GPT-Live-1 went generally available in the API yesterday, September 10, on the only endpoint it supports, v1/live/sessions, and on a billing unit OpenAI has never used for a conversational model before: $0.05 per minute of session, metered per second, no rounding up. That is $3.00 an hour. Three things sit underneath that number. The clock does not stop for silence, so a caller on hold costs the same as a caller talking. The model that answers the question is not the model you are paying $0.05 for; GPT-Live-1 handles the audio and delegates the thinking to a Responses model of your choice, and that model bills its normal token rates on top. And the rate itself is a strange piece of positioning. Five cents a minute is what Vapi charges to host a voice pipeline built from other people's models. OpenAI is now selling its own model into that slot, at that price, with the intelligence line still to come. We priced a 10-minute call four ways and got $0.51 to $0.86. The same call on gpt-realtime-2.1 came out between $0.51 and $3.22, and the reason for that spread is the more interesting story.
Image source: OpenAI
One line on the rate card, four lines on the invoice
Everything in this section is off OpenAI's own model page, pricing page and voice cost guide, all read on September 11. The model page says it in one sentence: "$0.05 per minute, billed per second," and the cost guide adds that "session duration is not rounded up to the next whole minute." No cached rate, no batch rate, no flex rate. The batch endpoint is listed as unsupported, along with every other endpoint except live sessions.
| Charge | Unit | Rate | Where it says so |
|---|---|---|---|
| Voice session | Second of active session | $0.05 per minute | Model page, pricing page |
| WebRTC session creation | 15 seconds, billed at init | $0.0125, credited once running | WebRTC guide, cost guide |
| Backend model (delegation) | Input and output tokens | That model's normal rate; 2x on priority | Delegation guide |
| Web search tool | 1,000 calls | $10.00 plus content tokens | Pricing page |
| Data residency endpoint | 10% uplift, post-March 2026 models | $0.055 per minute, our inference | Pricing page footnote |
The second row is the one nobody has written about. Creating a session over WebRTC, which is what a browser client does, bills 15 seconds of duration the moment the POST returns. It is credited against the session once it starts running, so a normal call never notices. A session that gets created and abandoned, or a client that reconnects in a loop, notices: 100,000 abandoned sessions is $1,250 for zero seconds of conversation. The WebSocket path carries no such note.
The last row is ours, not OpenAI's. The pricing page carries a footnote that regional endpoints charge a 10% uplift on models released on or after March 5, 2026, and the data controls page lists live sessions on the US and EU endpoints. Put those together and a residency-bound deployment pays $0.055 a minute. OpenAI has not printed a separate regional row for GPT-Live-1, so if that line matters to you, confirm it on your invoice rather than on this page.
The clock runs whether anyone is talking or not
This is the part of the cost guide I would pin above the desk of anyone building on it. Active session time, in OpenAI's own words, "includes time when the user speaks, the assistant speaks, both are silent, or the backend is working." Muting the microphone does not close the session. Three minutes of hold music is $0.15. A caller who wanders off to find their account number bills at the same rate as one mid-sentence.
It is worth saying plainly how different that is from the Realtime API, which sits one tab over in the same guide. Realtime bills audio tokens, and its voice activity detection filters out empty input, so silence costs approximately nothing. Two OpenAI voice products, one meter that ignores quiet and one that charges for it. The guide knows this, and its advice is unusually direct for a vendor. The cost guide says to close sessions promptly, because "if these optimizations help the user finish and close the voice session one minute sooner, you save $0.05 in voice charges." I have read a lot of pricing pages. Very few of them tell you to get the customer off the phone.
The other line in that sentence, "or the backend is working," is the one that ties the two bills together. When GPT-Live-1 delegates a question to the backend model, the voice clock keeps running while the backend reasons. A slow backend costs you twice: its own tokens, and the seconds of voice time spent waiting for them. The guide draws the conclusion itself, that "a larger backend model can cost less overall if it completes the task faster." That is true, and it is also the first time I have seen a vendor argue for its expensive model on the grounds that a cheaper one keeps a second meter running.
Per-minute against per-token, on OpenAI's own conversion
To compare a flat rate with a token rate you need a density, and OpenAI publishes one for Realtime: one audio token per 100ms of user speech, one per 50ms of assistant speech, so 600 input tokens and 1,200 output tokens per minute of the respective party talking. Google publishes 25 tokens a second for Gemini Live, and prints its own per-minute figures next to the token ones. Each vendor is converted with its own number below, never one vendor's ratio applied to another.
| One minute of | gpt-realtime-2.1 | gpt-realtime-2.1-mini | Gemini 3.1 Flash Live | GPT-Live-1 |
|---|---|---|---|---|
| User speaking | $0.0192 | $0.0060 | $0.0050 | $0.0500 |
| Model speaking | $0.0768 | $0.0240 | $0.0180 | $0.0500 |
| Half and half | $0.0480 | $0.0150 | $0.0115 | $0.0500 |
| Silence, or backend working | about $0 | about $0 | about $0 | $0.0500 |
| Conversation history re-sent per turn | $0.40/M cached, $32/M not | $0.30/M cached, $10/M not | Accumulates in context | None |
Read the half-and-half row and the two OpenAI products are essentially the same price, $0.048 against $0.050. That is not an accident and it is also not the comparison that matters. The last row is. Realtime resends the whole conversation as input on every response, so a 10-minute call with a turn every 30 seconds pushes roughly 85,500 tokens of history back through the input meter over its life. At the cached audio rate of $0.40 per million that is $0.03. At the uncached rate of $32 it is $2.74. OpenAI describes cache hits as best-effort and says any edit to the history busts them. So the same 10-minute Realtime call costs $0.51 with caching working and $3.22 with it not, and that 6.3x spread is the reason OpenAI's own cost guide carried a "typical" figure of around $0.30 a minute when we wrote up gpt-realtime-2 in May, six times the raw audio arithmetic.
GPT-Live-1 deletes that whole column. There is no history re-billing, because there is no per-token meter on the voice layer at all, and there is no cache discount to lose. What you get instead is a flat $0.50 for those 10 minutes and a different variable to worry about, which is the next section. If your Realtime cache hit rate has been good, Live is roughly a wash on the voice layer and you are paying for the backend on top. If it has been bad, Live is the cheaper product before the backend is even counted, and the hardest part of budgeting a voice agent just went away.
Where the backend lands, from one and a half percent to forty-two
GPT-Live-1 does not answer hard questions itself. It runs the conversation, and when the caller asks for something that needs thinking, it delegates to a Responses model, keeps talking while that model works, and folds the answer back in. The delegation guide recommends starting with GPT-5.6 Terra, or Luna "for cost-sensitive workloads," and the announcement names Astra "for complex customer issues." Only function and web_search tools are allowed on that path. No file search, no code interpreter, no MCP through the managed delegation.
The context each delegation carries is not published, so the table needs an assumption and here is ours: a 10-minute support call, eight delegations, each sending about 3,000 input tokens of instructions plus conversation and getting about 300 tokens back. That is 24,000 input and 2,400 output tokens for the whole call. Scale it to your own logs; the shape is what matters.
| Backend model | Rate per 1M in / out | Backend tokens | Voice, 10 min | Call total | Backend share |
|---|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 / $1.20 | $0.008 | $0.50 | $0.51 | 1.5% |
| GPT-5.6 Terra | $2.00 / $12.00 | $0.077 | $0.50 | $0.58 | 13.3% |
| GPT-5.6 Sol | $4.00 / $20.00 | $0.144 | $0.50 | $0.64 | 22.4% |
| GPT-6 Astra | $10.00 / $50.00 | $0.360 | $0.50 | $0.86 | 41.9% |
On Luna the backend is a rounding error and the call is the voice layer plus nothing. On Astra it is the better part of half the bill, and that is before priority processing, which the delegation config exposes as a service_tier and which doubles every backend token rate. OpenAI's own worked example on the cost guide, a 90-second session at $0.075 voice plus $0.02 backend, sits at about 21% backend, which lines up with a Terra-class model on a short call. The two rate cards you actually need for this are Luna to Astra on our pricing table, and none of them moved yesterday.
One thing the delegation model does for a budget that per-token voice never could: the expensive part is now a decision you make per request. Route order-status questions to Luna, route the angry cancellation to Astra, and the voice layer costs the same $0.05 either way. Client-side delegation goes further and lets the backend be "any model, agent, or service your application operates," including one that is not OpenAI's. That is a genuinely open door, and it is also why OpenAI can price the voice layer at the platform rate. It is competing for the pipe, not the brain.
Five cents is the price of the plumbing
Here is what jumped out when we lined the number up against the voice agent market. Vapi charges $0.05 a minute to host a pipeline and passes speech-to-text, the LLM and text-to-speech through at cost. Retell charges $0.055 for infrastructure and then adds the model and the voice as per-minute lines, with GPT 5.5 at $0.16 a minute on its card. Neither of those $0.05 figures includes a model. OpenAI has now put a model in at the same price, and made the model the optional extra.
| Product | Per minute | Per hour | What the number includes |
|---|---|---|---|
| Gemini 3.1 Flash Live, half and half | $0.0115 | $0.69 | Audio tokens only; text and context on top; preview |
| gpt-realtime-2.1, half and half | $0.048 | $2.88 | Audio tokens only; history re-billing on top |
| GPT-Live-1 | $0.05 | $3.00 | Voice model and session; backend tokens on top |
| Vapi | $0.05 | $3.00 | Hosting only; STT, LLM and TTS at cost |
| Retell | $0.055 | $3.30 | Infrastructure only; LLM $0.003 to $0.16/min and voices extra |
| ElevenLabs Agents | $0.08 | $4.80 | Speech engine; LLM and telephony billed separately |
| Grok Voice Think Fast 2.0 | $0.08 | $4.80 | Audio sent or received; $0.004 per text event |
So the honest per-hour range for a working GPT-Live-1 agent is about $3.05 on Luna to $5.16 on Astra, using the delegation profile above. That puts it under both $0.08 products even at the top of the range, and it puts it above Realtime only when Realtime's cache is behaving. It also does something xAI's meter, which we complained about in August, still does not: it defines the unit. Session duration, wall clock, per second. There is no ambiguity about whether a two-way minute meters once or twice, because it meters the session, not the audio.
The benchmarks, and the footnote about which backend ran them
The announcement's headline claim is a 30-point jump on Full Duplex Bench over gpt-realtime-2.1, and the charts on the post put the numbers at 80.1% against 45.4% on the v1.5 interactivity measure, with turn-taking latency at 0.798 seconds against 1.41. Full duplex is the point of the architecture: the model listens while it speaks, which is what lets it be interrupted without a chained STT-LLM-TTS handoff. Speak, one of the launch customers, reports interruptions cut by almost 80% against its previous turn-based system.
The agentic number is the one to read with the footnote. OpenAI says GPT-Live-1 ranks first on Tau3, at 86.2% pass@1 against 45.7% for gpt-realtime-2.1, and the chart is labelled "GPT Live backend: Astra (medium)." The tool-use charts are labelled Terra at low effort. Those are not the same product as the $0.05 line. They are the $0.05 line plus a $10 and $50 model, or plus a $2 and $12 one, and the score you get with Luna behind it is not published anywhere. On Artificial Analysis' conversational dynamics measure, which does not involve a backend, the gap is 97.3% to 95.7%, real but small. The big deltas are in the tasks where the backend does the work.
None of that is a criticism of the numbers, which are the numbers. It is a reminder that a benchmark chart for a voice front-end is really a chart of a two-model system, and the cost of reproducing the top row is the Astra column of the table above, not the $0.50 on the left.
Limits, and the things that are not on any page
Rate limits are concurrent sessions rather than tokens: 25 at Tier 1, 50 at Tier 2, 200 at Tier 3, 300 at Tier 4 and 500 at Tier 5, with the free tier unsupported. A Tier 1 account therefore cannot spend more than $75 an hour on the voice layer no matter how hard it tries, and a Tier 5 account tops out at $1,500 an hour. That is a very different ceiling from a token meter, and a much easier one to reason about.
Context is 128,000 tokens and it includes, in OpenAI's words, "audio tokens that don't appear in the transcript." Past 90% the session spins up a replacement voice engine with the original instructions and up to 8,192 tokens of history or summary, so a long call loses detail quietly rather than failing loudly. Instructions cap at 16,384 tokens. There is a session.closed reason of expired described as the session reaching its duration limit, and the duration limit is not published. We looked. There is also a GPT-Live-1 mini in ChatGPT, where it is the free tier's default, and no mini in the API; the model page URL for it returns a 404.
Twelve voices beyond the default, four of them marked generated, covering Australian, British, Irish, North American, Brazilian Portuguese, Filipino and Southern US English. Stored sessions, off by default, keep a recording for 30 days and are unavailable under zero data retention. Knowledge cutoff is July 31, 2025. Data residency covers the US and EU endpoints only.
And a note on the reaction, or the absence of one. The July ChatGPT launch of GPT-Live ran to 750 points and 527 comments on Hacker News. The API launch thread yesterday had 10 points and one comment at the time of writing, which asked for a single changelog page. Nobody has argued about the price yet. I suspect that is because $0.05 a minute reads as cheap, and it is, right up until someone sums a month of hold time.
Where this leaves a voice budget
If you are on Realtime and your cache hit rate is anything short of excellent, the arithmetic says move, and the flat rate will make your finance team happier than any benchmark will. If your cache is behaving and your calls are short and dense, Realtime stays a hair cheaper and you keep the silence discount, which matters more than it sounds for anything with a queue in front of it.
On the backend, start where the docs tell you to, on Terra, and then measure how many delegations actually need it. Our guess from the profile above is that most support traffic is Luna traffic with a few Astra moments, and routing per request is the entire game. Set max_output_tokens on the delegation, because a backend that rambles costs tokens and voice seconds at once. Close sessions the moment the caller is done, because OpenAI told you to and put a price on it.
And if you are choosing between this and a platform, the question has flipped. Vapi and Retell used to charge you $0.05 for the plumbing and let you choose the brain. OpenAI now charges you $0.05 for the plumbing plus a brain that is good at the plumbing, and lets you choose the second brain. Whether that is a better deal depends on how much of your bill was ever the model in the first place. For most voice agents we have seen the math on, it was not.
Sources
- OpenAI: gpt-live-1 model page - "$0.05 per minute, billed per second," the concurrent-session rate limits from 25 at Tier 1 to 500 at Tier 5, free tier unsupported, v1/live/sessions as the only supported endpoint, the statement that backend Responses calls use the normal pricing for the configured model and tools, audio and text in and out, July 31, 2025 knowledge cutoff
- OpenAI: API pricing - The GPT-Live-1 line, gpt-realtime-2.1 at $32.00 audio input, $0.40 cached and $64.00 audio output per million with the mini at $10.00, $0.30 and $20.00, the backend rate cards for Luna, Terra, Sol and Astra, the 2x priority tier, web search at $10.00 per 1,000 calls, and the footnote on the 10% regional uplift for models released on or after March 5, 2026
- OpenAI: voice latency and cost guide - The definition of active session time as including silence and backend work, the no-rounding rule, the 90-second worked example at $0.075 plus $0.02, the "save $0.05" advice, the "larger backend model can cost less overall" line, and on the Realtime tab the one-token-per-100ms and one-per-50ms audio density plus the note that VAD filters empty audio
- OpenAI: WebRTC guide for Live - Where session creation over WebRTC is documented as billing 15 seconds of voice duration at initialization, credited against the session once it runs
- OpenAI: Live delegation guide - The recommendation to start with Terra or use Luna for cost-sensitive work, the function and web_search tool restriction, the service_tier, reasoning and max_output_tokens knobs, and client delegation to "any model, agent, or service your application operates"
- OpenAI: Live conversations guide - The 128,000-token context including inaudible audio tokens, the replacement voice engine at 90% with up to 8,192 tokens of history, the 16,384-token instruction cap, the twelve additional voices, 30-day stored sessions, and the expired close reason without a published duration limit
- OpenAI: introducing GPT-Live-1 in the API - The September 10 announcement. Source for the 30-point Full Duplex Bench claim, the Tau3 first-place claim with its Astra (medium) backend footnote, the Terra (low) footnote on tool-use charts, the Speak interruption figure, and the Luna and Astra backend examples. Chart values of 80.1% and 45.4% interactivity, 0.798s and 1.41s latency, 86.2% and 45.7% Tau3 and 97.3% and 95.7% conversational dynamics were read from the post's charts; the page returns 403 to a direct fetch and was read through a text-extraction proxy and secondary coverage
- OpenAI: API changelog - The September 10 entry stating GPT-Live 1 is generally available in the API
- OpenAI: data controls and residency - v1/live/sessions listed on the United States and European regional endpoints only, and stored sessions unavailable under zero data retention
- Google: Gemini API pricing - gemini-3.1-flash-live-preview at $3.00 per million or $0.005 per minute audio input and $12.00 per million or $0.018 per minute audio output, page dated September 8, 2026, and the 25 tokens per second conversion
- xAI: models and pricing - grok-voice-think-fast-2.0 at $0.08 per minute or $4.80 per hour of audio, plus $0.004 per text input
- ElevenLabs: Agents pricing - $0.08 per minute with the LLM and telephony billed separately on top, and $0.16 per minute burst pricing past the concurrency limit
- Vapi: pricing - $0.05 per minute hosting with speech-to-text, LLM and text-to-speech passed through at cost, or free with your own keys
- Retell: pricing - Voice infrastructure at $0.055 per minute, LLM add-ons from $0.003 per minute for GPT 5 nano to $0.16 for GPT 5.5 on the standard tier and double that on fast, voices at $0.015 to $0.040 per minute, and telephony at $0.015 per minute in the US
- Hacker News: GPT-Live-1 in the API - The API launch thread, 10 points and one comment at the time of writing, against 750 points and 527 comments on the July ChatGPT launch thread