What GPT-Live 1.1 Charges While a Voice Call Is Silent OpenAI released GPT-Live-1.1 in its API on November 19 and added speech-triggered standby for WebRTC sessions on both GPT-Live generations, letting applications pause inference while a call connection stays open. At OpenAI's published rates, an active voice minute costs $0.05 and a standby minute costs $0.02, so a ten-minute call with six active and four standby minutes costs $0.38 in voice charges versus $0.50 if the model stays active throughout — a $0.12 saving, or 24% of the all-active voice charge. OpenAI meters duration by the second, and a team can keep its existing GPT-Live-1 model and adopt the new billing behavior without a model upgrade. A customer calls about a replacement part. The agent asks for the serial number. “Give me a minute. It’s on the back of the machine.” The customer sets the phone down. On a ten-minute call, keeping GPT-Live-1.1 active the whole time costs $0.50 in voice charges. If four of those minutes are actually spent in standby, the voice charge falls to $0.38. Both budgets buy ten minutes of connection. The difference is what the service is doing while the customer looks. On November 19, OpenAI released GPT-Live-1.1 in its API and added speech-triggered standby for WebRTC sessions on both GPT-Live generations. WebRTC is the connection a browser or app can use to carry the conversation. While that connection stays open, the application can let the model pause and wake it with the caller’s voice. A team can keep its existing GPT-Live-1 model and try the new billing behavior independently of a model upgrade. OpenAI’s API changelog https://developers.openai.com/api/docs/changelog 1. 01Ten minutes can carry two ratesAt the published rates, six active minutes and four standby minutes cost $0.38 in voice charges. Ten active minutes cost $0.50. 2. 02Wait for the pause eventA caller can be silent while inference continues. The interval before a configured idle timeout still accrues active charges. 3. 03Keep backend work in the estimateA stock lookup or reasoning call can cost money alongside either voice rate. Make the budget cover the full completed request. 4. 04Give a returning caller a way backSpeech can wake a standby session. A closed session needs a new connection and a way to restore the case and its outstanding work. 01 — Voice minutesWhat makes a quiet minute billable? OpenAI quotes two prices for GPT-Live-1.1. Each active minute costs $0.05. Each standby minute costs $0.02. The service meters duration by the second, so a sequence of short pauses can matter on a busy support line. Standby means the model has paused inference, its work of processing the conversation. An open connection with two quiet people on it can still be running inference at the active rate. OpenAI’s model pricing https://developers.openai.com/api/docs/models/gpt-live-1.1 Consider two callers who each spend a minute searching for a label. The first keeps asking which number to read while the assistant describes where to find it. The second puts the phone down and searches without speaking. Those calls have the same purpose and wall-clock duration, yet only the second is an obvious candidate for sustained standby. The first customer is using that minute of assistance. Cutting it out of a budget simply because the support script calls the step a “hold” would understate the cost. The price covers the voice session. A reasoning model that decides which replacement fits the appliance can accrue separate charges, as can a tool that looks up stock. Keep the voice line, the backend model and any paid lookup on separate rows in the estimate. A manager can then tell whether a change helped customers find the serial number more quickly or simply reduced the price of keeping the line open. 02 — Worked exampleTwo budgets for the same ten minutes Suppose the call takes ten minutes from start to finish. In the first version the voice model stays active throughout. In the second, six minutes are billed as active and four as standby. Six active minutes at five cents contribute $0.30. Four standby minutes at two cents contribute $0.08. The $0.12 saving is 24% of the $0.50 voice charge in the first version. It is a smaller percentage of the total bill whenever both versions also incur the same backend charges. The third line below accounts for a detail that changes the result. Imagine four uninterrupted minutes of quiet with a 30-second timeout after the last transcript. On the simplified assumption that inference pauses when that timer expires and resumes as the quiet period ends, only three and a half minutes qualify. The first half minute adds $0.025 at the active rate. Moving that half minute out of the original standby budget raises the total by $0.015 because it also removes $0.01 of standby time. The amount shown is an unrounded estimate, before any invoice aggregation. For a volume forecast, multiply by calls that resemble this one. A thousand calls with four actual standby minutes each would save $120 in voice charges against leaving all thousand active. A thousand calls on which someone occasionally pauses for breath is a different workload. Sample the calls where customers leave the conversation open, then use their share of traffic when building the monthly forecast. A support director may have a more immediate question: what happens to money already set aside for a pilot? Suppose a queue expects 100 calls like the second example and sets aside $50 for the voice portion. At $0.38 each it would use $38, leaving $12 of that allocation. If a backend lookup on every call costs another $0.04 in this hypothetical budget, total modeled spend becomes $42. Compare that with $54 for the same backend work on 100 always-active calls. The saving is still $12. The percentage has changed because the denominator now includes the lookup. 03 — Session settingsWhen does the clock switch rates? When creating the WebRTC session, set session.wake policy.type to speech . That tells the service to wait for speech before starting inference. Set session.idle timeout s to a whole number from 5 through 60 to have it return to standby after that many seconds without an input or output transcript. If the timeout is omitted, inference remains active after its first wake for the rest of that session. A speech-based start alone therefore does little for a long hold halfway through the call. OpenAI’s standby configuration https://developers.openai.com/api/docs/guides/live-conversations configure-standby A timeout applies to each quiet interval. Picture four one-minute pauses and a 30-second timeout, with a spoken exchange between pauses. Assuming the pauses and wakes line up with that pattern, each interval contributes about 30 seconds of standby. Two minutes qualify across the call. A single uninterrupted four-minute pause under the same assumptions contributes three and a half. A recording can contain the same amount of silence in each case and yield a different voice bill. Keep the microphone track active when the caller expects speech to wake the conversation. Muting it is a separate user action. A mute control should describe what happens to the microphone, and a hold control should describe how to return to the call. Giving the two controls the same label invites a caller to start reading out a serial number while the app is still waiting for microphone input. 04 — Caller experienceWhat the customer hears on returning During standby the connection and session ID remain in place. OpenAI restores the conversation as the caller speaks. Its guide warns that the first response to a short utterance after standby can take longer. A customer who returns with “Found it” may notice that pause more than someone who immediately reads a long serial number. No published delay in milliseconds is needed to design a test for that difference. Try two return phrases with people who have the same information in front of them. One says “Hello?” and waits. Another starts reading the number before the agent acknowledges them. Check the recording and transcript for the opening characters, time the end of the customer’s speech to the first useful reply, and note whether either person repeats themselves. A prompt acknowledgement followed by a request to repeat the number still costs the customer a turn. The voice latency measurement guide https://www.digitalapplied.com/blog/voice-agent-latency-measures-reference explains why first audio and a useful answer deserve separate timestamps. Choose the timeout with that return experience in mind. A short timer favors a service desk that regularly sends people away to fetch information. On a call where a customer dictates an address, they may stop to spell a street or find a postal code several times. A longer timer can keep those exchanges flowing. Replay that specific pattern before extending an aggressive standby setting to every queue. 05 — Pending workWho pays while the tool is looking something up? The customer reads the number. The voice agent hands it to a parts lookup. For the next 20 seconds the customer waits for stock availability. The application might show a spinner for the full 20 seconds, but that tells finance nothing about the voice rate. OpenAI counts time waiting for a backend operation at the active rate while voice inference is running. If an eligible session actually pauses during a wait, the standby interval is billed at the lower rate. Meanwhile the tool or reasoning model has its own meter. OpenAI’s cost guide https://developers.openai.com/api/docs/guides/voice-latency-cost?api=live Keep the stock check under an application task ID so that a delayed reply can be matched to the current request. A caller might correct the model number while the first lookup is running. Before speaking, the application needs to know which lookup finished, whether it succeeded, and whether its result still answers the question. A pause in voice activity is poor evidence for any of those facts. The guide to background work on voice calls https://www.digitalapplied.com/blog/voice-agent-background-tool-work-guide follows that handoff in more detail. If finance asks why an expensive call was quiet for most of its duration, reconstruct both timelines. Mark when the customer spoke, when the tool started and returned, and when the voice session paused and resumed. A slow lookup with continuous progress announcements suggests a different fix from a caller spending five minutes walking to another room. Shortening the query helps the first caller. Keeping the second caller connected at the cheaper rate preserves the option of asking for help. There is a simple way to judge a proposal for a faster backend. Suppose a quicker lookup costs an extra cent but reliably takes 30 seconds off a call and the customer can hang up immediately afterward. At five cents an active minute, avoiding those 30 active seconds saves two and a half cents in voice charges, leaving one and a half cents of net savings in this illustration. If all 30 seconds would have been billed in standby, they cost one cent in the first place. The upgrade then breaks even on this comparison. A customer may still value a faster answer. Write that down as a benefit alongside the arithmetic. 06 — Longer absencesWhen is it better to end the session? A caller who has to find a paper receipt upstairs may want to return by voice. Someone waiting until a warehouse reopens tomorrow needs a way back to the case, with no good reason to keep a microphone listening in the meantime. Offer a visible or spoken way to end the conversation, preserve the outstanding task in the application, and provide a clear way to resume. A closed Live session cannot hear someone say they have returned. Before closing, finish work that depends on the voice session and let any promised spoken message play. OpenAI documents a graceful close in which the application installs a listener, requests closure and waits for the final event before releasing the connection. Save the case identifier, outstanding operation and verified result as appropriate for the support service. When the caller returns, the new session can pick up with a relevant question. Avoid submitting a second replacement order just because a new voice session has a new ID. Repeated disconnections also have a startup cost. OpenAI’s cost guide says that under the default WebRTC startup behavior, session creation initially bills 15 seconds which are credited against the duration once the session starts. A running 90-second session already includes those seconds. Adding a fresh 15 seconds to every completed call would overstate the forecast. Consider both session creation and the disruption of reconnecting when deciding how long a caller can be away. OpenAI’s WebRTC initialization charges https://developers.openai.com/api/docs/guides/voice-latency-cost?api=live webrtc-initialization-charges Think through the first sentence of the reopened call. A customer who comes back tomorrow wants to hear whether the part is available. Asking for the serial number again defeats much of the value of resuming. Before reading back a saved case, let the support application establish that the caller is entitled to it, recover the actual warehouse result and give the voice model just the information needed for that exchange. A saved transcript alone describes what people said yesterday. The application record can establish whether the warehouse ever confirmed stock or whether a different employee has already resolved the question. Spelling a street name Keep voice ready through the hesitation between a house number and the spelling of a street. Finding the appliance label Maintain the WebRTC connection so the person can read the serial number aloud when they find it. Waiting for the warehouse Save the support case and provide a way to restart once a person or system has the answer. 07 — Pilot designReconcile the call before setting a default During a small rollout, listen for session.paused and session.resumed and retain their timing against the call identifier. The resumed event can also mark the first wake. Keep the latest cumulative voice usage for the call and collect the final usage when it closes. Summing every running update as though it were a new interval would count the same opening minutes several times. Treat inferred state intervals as an operational estimate to reconcile with usage records. Keep backend usage from delegated responses alongside it. Use three ordinary customer journeys for comparison. One customer reads the requested information without leaving. Another returns after a long search. A third waits for the backend and changes part of the request. For each, record the voice spend, backend spend, the customer’s first useful reply after a pause and whether the correct operation completed. Time from wake to playback has more meaning alongside the customer’s words. A caller who says “Are you there?” twice has already told the team something that a split-second server metric cannot. The outcome to compare is cost per resolved case. If one queue saves twelve cents on a call but more people have to call back with the same question, include those repeat calls. If a long pause used to cause customers to hang up, an inexpensive open line can be worth paying for even when a disconnect would technically cost less. The business guide to GPT-Live https://www.digitalapplied.com/blog/gpt-live-1-api-business-voice-guide explains how to place model delegation and the completed customer task on the same report. For a buyer comparing vendors, take the same customer journey to each pricing sheet. A provider that quotes audio minutes, audio tokens, or separate input and output prices may count the quiet interval differently. A headline comparison using ten times each service’s per-minute number can miss that distinction altogether. Ask for the billing units generated by the sample, then add backend work on the same terms. The comparison of Microsoft’s voice pricing https://www.digitalapplied.com/blog/microsoft-mai-realtime-1-voice-ai-price illustrates why even the words “audio minute” require a precise definition before the numbers belong in one procurement table. For a team building this into a broader support workflow, our AI transformation service https://www.digitalapplied.com/services/ai-transformation covers integration with the systems that actually resolve the case. Pair the final call usage with the outcome of the parts lookup. A complete transcript still leaves open the most important question for the customer: did the correct replacement get reserved? Give each support queue a reason to stay connected Set the standby policy where the work happens. Keep a fast exchange active, use speech-triggered standby while someone looks for information, and close the line when the next useful step is hours away. The customer should always know how to come back. Finance should be able to explain the charge for that exact call.