Quick answer: OpenAI owns more of a phone call’s path: its Realtime API and GPT-Live accept 8 kHz G.711 μ-law and terminate SIP; Gemini 3.8 Live takes PCM over a WebSocket and returns 24 kHz audio you must re-encode. A 5-minute call cost $0.29–$0.43 on Gemini 3.8 Live, $0.25 plus backend on GPT-Live 1, and $0.14–$1.87 on OpenAI’s token-billed models.
Every figure here was read on the vendor’s own pages on 26 September 2026, and the line-up moved this month. Google made gemini-3.8-live generally available in the Gemini API on 15 September 2026 (release notes), and its Agent Platform model page gives a release date of 24 September 2026; OpenAI made gpt-live-1 generally available on 10 September 2026 (changelog), beside the token-billed Realtime API rather than replacing it. So this page compares three products: Gemini 3.8 Live, OpenAI gpt-realtime-2.1 and OpenAI gpt-live-1. Google prices in USD; OpenAI’s page prints a bare “$”, reproduced as printed.
Gemini Live API vs OpenAI Realtime API: which is better for a phone voice agent?
It depends on which seam you want to own. If your calls arrive on a SIP trunk or a Twilio Media Stream, which is “always” G.711 μ-law at 8 kHz, OpenAI is the shorter path: both its voice APIs accept that format and can terminate SIP themselves. Google’s Live API takes raw 16-bit PCM over a WebSocket and returns 24 kHz PCM, so a phone deployment needs a bridge that you or a partner runs.
On the price of a fresh minute of audio, Google is lower: $3.00 and $12.00 per million audio input and output tokens, against $32.00 and $64.00 for gpt-realtime-2.1. But both vendors re-bill the conversation so far on every turn, only OpenAI caches it, and gpt-live-1 ignores tokens and bills $0.05 a minute. For a phone agent, OpenAI wins the telephony seam, Google wins the per-token rate, and GPT-Live 1 is the only one of the three whose voice bill you can forecast from call length alone.
The Three-Seam Test: the criteria that decide it on a phone line
A browser demo never touches them, but a phone line adds three seams. Checking them is what we call the Three-Seam Test: before comparing voices, ask who owns each seam on each API.
- The codec seam. The phone network carries G.711 at 8 kHz. Does the model accept it as sent, or must something transcode it both ways inside your latency budget?
- The session seam. Does one model session outlast a 20-minute call, and what survives a mid-call reconnect?
- The signalling seam. Who answers the SIP INVITE, places the outbound call and sends the REFER to a human?
Cost, languages and residency come after those, because a seam you own is engineering time on every release.
The comparison table, telephony criteria first
| Criterion | Google Gemini 3.8 Live (Live API) | OpenAI gpt-realtime-2.1 (Realtime API) | OpenAI gpt-live-1 (Live API) |
|---|---|---|---|
| Transport | “Stateful WebSocket connection (WSS)” (overview) | WebRTC, WebSocket or SIP (SIP guide) | WebRTC, WebSocket or SIP (SIP guide) |
| Audio in | Raw 16-bit PCM, natively 16 kHz; “will resample if needed so any sample rate can be sent” (capabilities) | PCM at 24 kHz only, G.711 μ-law or G.711 A-law (API reference) | PCM16 at 24 kHz (default) or 16 kHz, G.711 μ-law or A-law at 8 kHz; one format for both directions (WebSocket guide) |
| Audio out | Raw 16-bit PCM, “always” 24 kHz | Same format list as input, set separately | Same format as input |
| Native SIP, inbound | None on the 15 Google pages we read; telephony via partners (Voximplant, Twilio, Daily, LiveKit) | Yes: sip.api.openai.com, EU sip-eu.api.openai.com, TLS |
Yes, once SIP support is enabled for the project; TLS signalling, SRTP audio required |
| Native SIP, outbound | None on the pages we read | No: “not the Realtime API call-creation endpoint” | Yes, if enabled for your organisation; trunk needs TLS, Opus, SDES-SRTP; E.164 destinations only |
| Session length | Connection “around 10 minutes”; audio-only session 15 minutes without compression; resumption token valid 2 hours (session management) | “The maximum duration of a Realtime session is 60 minutes” (conversations guide) | Outbound SIP connected call limited to 2 hours; context engine replaced at 90% of 128,000 tokens (sessions guide) |
| Context window | 131,072 input tokens (model page) | 128,000 tokens (model page) | 128,000 tokens by default |
| Function calls while speaking | Async (NON_BLOCKING) by default, with SILENT, WHEN_IDLE and INTERRUPTED scheduling | The voice model selects the functions itself | Delegated to a backend model or your own agent; GPT-Live “can continue speaking while the backend works” |
| Turn detection | Automatic VAD with sensitivity, padding and silence settings; can be disabled; hybrid VAD | server_vad or semantic_vad; can be turned off (VAD guide) |
Full duplex; “GPT-Live decides when to speak” (migration guide) |
| Audio price | $3.00 in / $12.00 out per 1M tokens; 25 tokens per second (Agent Platform pricing) | $32.00 in / $0.40 cached / $64.00 out per 1M tokens (pricing) | $0.05 per minute, billed per second, backend extra |
| Prompt caching | “Not supported” (model page) | Automatic, “best-effort and not guaranteed” (cost guide) | Not applicable to the voice session; applies to backend models |
| Regional processing | Agent Platform: us and eu multi-regions, us-central1 (model page) |
United States, Europe (EEA + Switzerland) (Your data) | United States, Europe (EEA + Switzerland) |
| Languages | 70 (overview), 99 listed (capabilities table), 24 (Agent Platform page) | No count on the pages we read | No count; 12 additional voices listed by regional influence, two Australian English |
| Shutdown date | “No shutdown date announced” (deprecations) | Current replacement for the gpt-realtime family, which shuts down 20 January 2027 (deprecations) |
Generally available since 10 September 2026 |
OpenAI owns all three phone seams on GPT-Live and all but outbound signalling on the Realtime API; Google’s Live API leaves all three with you or a partner. That is the whole telephony argument in one line, and everything below is the detail behind it.
Who resamples the audio: 8 kHz μ-law on each API
A caller reaches you as G.711 μ-law: 8,000 one-byte samples a second. Twilio’s Media Streams documentation says the encoding is always audio/x-mulaw at 8000 Hz on one channel.
On OpenAI, nothing has to happen. The Realtime API reference lists PCM (“Only a 24kHz sample rate is supported”), G.711 μ-law and G.711 A-law. The GPT-Live WebSocket guide adds 16 kHz PCM and says “A matching G.711 stream can pass through without conversion to PCM.” Over native SIP you set no format at all: “SIP negotiates the audio format, so omit audio.format.”
On Google, your bridge does the work both ways. Audio “is always raw, little-endian, 16-bit PCM”. The Live API will resample whatever input rate you declare, so audio/pcm;rate=8000 is accepted, but you must decode μ-law to 16-bit samples first, doubling the stream to 16,000 bytes a second. Coming out, “Audio output always uses a sample rate of 24kHz”: 48,000 bytes a second that you downsample by three and re-encode to μ-law before the carrier will play it.
That is routine code, but it sits in your audio path, adds buffering that barge-in must account for, and is the first suspect when callers say the agent sounds muffled. On a phone line, Gemini Live adds a decode and a resample on the way in, and a resample and an encode on the way out, on every call.
How long can one phone call last on each API?
Google’s session management guide sets two limits. The session: “Without compression, audio-only sessions are limited to 15 minutes”, extendable by context window compression. The connection: “The lifetime of a connection is limited as well, to around 10 minutes. When the connection terminates, the session terminates as well.” The documented fix is session resumption, and “Resumption tokens are valid for 2 hr after the last sessions termination.”
So a 25-minute Gemini call is several WebSocket connections stitched together by your code. If that stitching is missing, the symptom is not a dropped call but an agent that greets the caller again, which we traced on why a voice agent loses context mid-call after a reconnect.
OpenAI states one ceiling for the Realtime API: “The maximum duration of a Realtime session is 60 minutes.” For GPT-Live, the sessions guide lists expired (“The session reached its duration limit”) as a close reason, and the only number we found is in the SIP guide: “Ringing is limited to 3 minutes, and a connected call is limited to 2 hours” for outbound SIP. Long GPT-Live calls meet a context limit first: past 90% of the 128,000-token window it “starts a replacement voice engine within the same session” carrying your instructions and up to 8,192 tokens of history.
What does a 5-minute phone call cost on each?
None of the three has a per-call price you can look up, so we price one call shape end to end on each, reusing the shape from our voice AI cost-per-minute breakdown of realtime versus pipeline so the numbers can be cross-checked. Our assumptions, not vendor claims:
- A 5-minute call of 20 turns. Each turn is 15 seconds: the caller speaks for 6, the agent replies for 9.
- 800 tokens of text instructions, sent on every turn.
- OpenAI audio tokens per its cost guide: “1 token per 100 ms” of user audio and “1 token per 50ms” of assistant audio, so 60 caller and 180 agent tokens per turn, plus 59 text tokens of agent transcript.
- Google audio tokens per its Agent Platform pricing, “25 tokens /second of audio”: 150 caller and 225 agent tokens per turn.
- Both vendors re-bill accumulated context. OpenAI: “The entire conversation is sent to the model for each Response.” Google: “Tokens from past turns are re-processed and billed in every new turn.”
- For GPT-Live, six backend delegations on
gpt-6-luna(OpenAI’s suggested starting backend) at 3,000 input and 150 output tokens each.
Every step for the first and last turn:
| Step | gpt-realtime-2.1, cache misses | Gemini 3.8 Live |
|---|---|---|
| Turn 1 text input | 800 × $4.00/1M = $0.00320 | 800 × $0.75/1M = $0.00060 |
| Turn 1 audio input | 60 × $32.00/1M = $0.00192 | 150 × $3.00/1M = $0.00045 |
| Turn 1 output | 180 × $64.00/1M + 59 × $24.00/1M = $0.01294 | 225 × $12.00/1M = $0.00270 |
| Turn 1 total | $0.01806 | $0.00375 |
| Turn 20 input carried | 1,921 text + 4,620 audio tokens | 800 text + 7,275 audio tokens |
| Turn 20 total | $0.00768 + $0.14784 + $0.01294 = $0.16846 | $0.00060 + $0.02183 + $0.00270 = $0.02513 |
| 20 turns: input summed | 27,210 text + 46,800 audio tokens | 16,000 text + 74,250 audio tokens |
| 20 turns: total | $0.10884 + $1.49760 + $0.25872 = $1.8652 | $0.01200 + $0.22275 + $0.05400 = $0.2888 |
Whole-call results:
| Product and scenario | 5-minute call, model cost | How it is billed |
|---|---|---|
| gpt-realtime-2.1-mini, cache holds | $0.1364 | Tokens, re-billed each turn, cached input at $0.30/1M audio |
| gpt-live-1, voice plus six gpt-6-luna delegations | $0.2500 + $0.0023 = $0.2523 | Seconds of active session, including silence |
| Gemini 3.8 Live, caller speech billed as input | $0.2888 | Tokens, re-billed each turn, no caching |
| Gemini 3.8 Live, all 15 seconds of line audio billed as input | $0.4305 | Same, with input counted while listening |
| gpt-realtime-2.1, cache holds | $0.4412 | Tokens, cached input at $0.40/1M |
| gpt-realtime-2.1-mini, cache misses | $0.5592 | Tokens, full rate |
| gpt-realtime-2.1, cache misses | $1.8652 | Tokens, full rate |
Gemini has two rows because Google’s pages leave a billing question open. The Agent Platform pricing page says that in Proactive Audio Mode “input tokens are charged while LiveAPI is listening”, but that footnote sits in its Gemini 2.5 section and the Gemini 3.8 Live billing notes do not repeat it; the Gemini 3.8 Live model page says “Proactive audio is now permanently enabled”. A phone bridge streams line audio for the whole call, so we price both readings rather than pick one. Ask Google which applies before you budget. Google’s printed per-minute equivalents ($0.005 in, $0.018 out) cover fresh audio only: $0.064 for this call, under a quarter of the re-billed $0.2888. And GPT-Live bills “time when the user speaks, the assistant speaks, both are silent, or the backend is working”, so hold time is not free.
Call length widens the gap between token billing and time billing. At 5 minutes Gemini 3.8 Live costs 14% more than GPT-Live; at 10 minutes on the same shape it costs $1.0275 against GPT-Live’s $0.50 plus about $0.0045 of backend, because re-billed context grows with the square of the turn count and a per-second rate does not. The carrier leg is extra on every row.
Can the agent call a function while it keeps talking?
On a phone call, two seconds of silence during a CRM lookup sounds like a dropped line. The three products answer that differently.
Gemini 3.8 Live runs function calls asynchronously by default. The model page makes asynchronous NON_BLOCKING execution “the default function calling mode” and says “Function scheduling (SILENT, WHEN_IDLE, INTERRUPTED) is supported”, which decides whether the model speaks about a result at once, after its current utterance, or silently later. Two cautions: the tools guide says “the Live API doesn’t support automatic tool response handling”, so your code returns every result, and its sample spells the value INTERRUPT where the model page says INTERRUPTED, so test what your SDK accepts. The previous model, Gemini 3.1 Flash Live, was sequential only.
gpt-realtime-2.1 selects and calls functions itself, the arrangement OpenAI’s migration guide describes as “Before: the Realtime model handles speech and selects functions”.
gpt-live-1 splits the job in two. The voice model decides when to delegate; a backend reasons and calls tools, either a Responses model OpenAI runs or “any model, agent, or service your application operates”. GPT-Live “can continue speaking while the backend works”, and your prompt splits in two as well.
Interruption and turn-taking: how much control do I get?
Phone audio is noisy, so turn-detection controls matter more than in a demo.
Gemini exposes the most: start and end sensitivity, prefixPaddingMs and silenceDurationMs (Google recommends 500 to 800 ms and says “The server’s internal default is approximately 800ms”), a manual mode with your own activityStart and activityEnd, and a hybrid. On barge-in, Google says the ongoing generation is dropped and the server “discards any pending function calls”, which your bridge must reconcile with work already under way.
The Realtime API offers server_vad, which chunks on silence with a threshold and padding, and semantic_vad, which uses “a semantic classifier to detect when the user has finished speaking” with a tunable eagerness. Either can be turned off.
GPT-Live removes the knob by design. The migration guide says to “remove manual audio commits and voice-turn triggers” because “GPT-Live decides when to speak”, and there is “no corresponding event marking the end of each spoken response”, so your player tracks when the agent stops. For exact wording, the sessions guide suggests playing “a verified recording or rendered clip” and controlling GPT-Live playback while it plays.
Languages, voices and where the audio is processed
Google published three language counts for this API on the day we read it: “Converse in 70 supported languages” on the Live API overview, a table of 99 languages in the capabilities guide, and “Converse in 24 supported languages” on the Agent Platform page. We found no language count for either OpenAI voice product; the GPT-Live sessions guide lists 12 additional voices by regional influence, two of them Australian English (Quartz and Ripple).
For regional processing, OpenAI’s Your data guide lists /v1/live/sessions and /v1/realtime in the United States and Europe (EEA and Switzerland); its Australian region is storage only, and the pricing page adds a 10% uplift on regional processing for eligible models released on or after 5 March 2026. Google’s Agent Platform lists gemini-3.8-live in the us and eu multi-regions and us-central1. Neither offered onshore Australian processing for these models on 26 September 2026. And on Google’s developer pricing page, the free tier’s “Used to improve our products” is “Yes”: keep real callers off free-tier keys.
What breaks in production: open issues as at 26 September 2026
Framework issue trackers show what documentation does not. These are the reports we would turn into test cases, each status observed on 26 September 2026.
- pipecat-ai/pipecat #5465, opened 27 August 2026: during a reconnect, Pipecat’s Gemini Live service drops tool results without logging, so “the model sits waiting on a call that never gets answered, so the bot just goes quiet for the rest of the turn”. Their turns ran “anywhere from 100s to ~1100s” on
gemini-3.1-flash-live-preview. Open as at 26 September 2026; linked fix PR #5497 unmerged as at the same date. - livekit/agents #5096, opened 12 March 2026: an end-call tool on
gemini-live-2.5-flash-native-audiovia Vertex AI “works ~70% of the time”, with a tool-reply timeout warning “95% of the times”. Open as at 26 September 2026, with linked PR #7186 open as at the same date. - livekit/agents PR #7390, opened 21 September 2026: describes Gemini Live in audio mode occasionally closing a turn with text and no audio, after which “the agent then sat silent until the caller hung up”. Open and unmerged as at 26 September 2026.
- On the OpenAI side, a developer forum report of 14 September 2026 says Realtime SIP
referandhangupreturned HTTP 200 while no REFER or BYE reached a Twilio Elastic SIP Trunk; one post and no replies as at 26 September 2026.
All three Gemini reports concern earlier Live models; gemini-3.8-live had been generally available for 11 days in the Gemini API and 2 days on the Agent Platform. Framework support for GPT-Live is equally new: Pipecat PR #5688, adding an OpenAI Live service for gpt-live-1, merged on 10 September 2026, and OpenAI’s partner integrations page warns that “A Realtime integration is not automatically compatible with GPT-Live.” The failure to test for on a phone agent is silence, and silence produces no error. Test a transfer, a hang-up, a slow tool call and a reconnect on every release.
Who each option is wrong for
Gemini 3.8 Live is wrong for you if no one on your team can own a WebSocket audio bridge, if you want the model vendor to handle SIP, or if your calls run past 10 minutes and resumption is not yet built.
gpt-realtime-2.1 is wrong for you if your instructions change mid-call. The cost guide says changing instructions or tool definitions mid-session “will reduce the cache rate for subsequent turns”, and on our call that is the difference between $0.44 and $1.87. It also cannot place outbound SIP calls; OpenAI routes those through the Live API.
gpt-live-1 is wrong for you if you need manual turn control, if you want the API rather than your own player to hold speech until a check completes (OpenAI puts such checks “in the audio playback path your application controls”), or if you need more than 25 concurrent calls on Tier 1; the model page lists 25, 50, 200, 300 and 500 for Tiers 1 to 5.
Whether a speech-to-speech model belongs in the call path at all is a separate question, weighed on speech-to-speech or pipeline, and the deprecation risk each carries. On notice periods: OpenAI told developers on 20 July 2026 that the gpt-realtime family shuts down on 20 January 2027, 184 days later; the most recent Live shutdown in Google’s release notes, gemini-2.0-flash-live-001, was announced on 20 October 2025 for 9 December 2025, 50 days later, on a model Google had released as a public preview. OpenAI’s most recent removal of a realtime snapshot, gpt-realtime-mini-2025-10-06, was announced on 22 April 2026 for 23 July 2026, 92 days later. If your model already has a date, see our migration plan for a deprecated voice model.
Build the bridge yourself, or hand the call path over?
Everything above is buildable. The question is whether the seams are where you want engineering time to go. Our thresholds, each tied to a published limit above:
| If this is true | Do this |
|---|---|
| You already run Twilio Media Streams, LiveKit or Pipecat and have an engineer on call for the audio path | Build directly on the API. The partner integrations exist for exactly this. |
| Your calls are under 10 minutes and inbound only | Any of the three works; choose on price and voice after a recorded test. |
| Your calls regularly pass 10 minutes | On Gemini, budget for resumption and GoAway handling before launch; on OpenAI, check the 60-minute Realtime ceiling. |
| You need outbound dialling from the model vendor over SIP | GPT-Live, after outbound SIP is enabled for your organisation, or any of the three behind your own carrier. |
| You need more than 25 concurrent calls on day one | Check your OpenAI tier before choosing GPT-Live; Tier 1 allows 25. |
| You need phone, SMS, email and WhatsApp follow-up on the same lead, in more than one language | The voice model is one layer of four. Price the orchestration, not just the minutes. |
That last row is the work Zian AI does. Zian runs autonomous sales agents across phone, SMS, email and WhatsApp, with SmartReach AI™ deciding the channel, message and timing of each follow-up and private model deployment on customer infrastructure where it is needed. Zian has been running outbound acquisition since 2017 and works across more than 10,000 leads a day. If you would rather own the leads than the audio bridge, Apply For Partnership while Zian is in partnership-application beta.
Where every figure on this page comes from
| Figure | Who published it | Link | Date read |
|---|---|---|---|
| gemini-3.8-live generally available 15 Sep 2026; 2.0 Live shutdown announced 20 Oct 2025 for 9 Dec 2025; 2.0 Flash Live released as a public preview | Gemini API release notes | 26 Sep 2026 | |
| gemini-3.8-live “No shutdown date announced” | Gemini deprecations | 26 Sep 2026 | |
| WebSocket protocol; 16 kHz PCM in, 24 kHz PCM out; 70 languages; partner list incl. Voximplant | Live API overview | 26 Sep 2026 | |
| Input resampled from any rate; output always 24 kHz; VAD defaults (~800 ms, 500–800 ms recommended); 99-language table; interruption discards pending calls | Live API capabilities guide | 26 Sep 2026 | |
| 15-minute audio-only session; ~10-minute connection; resumption tokens valid 2 hours | Session management with Live API | 26 Sep 2026 | |
| No automatic tool response handling; scheduling value INTERRUPT in sample; 3.1 Flash Live sequential only | Tool use with Live API | 26 Sep 2026 | |
| 131,072 input tokens; caching not supported; async default; SILENT, WHEN_IDLE, INTERRUPTED; proactive audio permanently enabled | Gemini 3.8 Live model page | 26 Sep 2026 | |
| $3.00/$12.00 audio, $0.75/$4.50 text per 1M; $0.005/min and $0.018/min equivalents; free tier used to improve products | Gemini Developer API pricing | 26 Sep 2026 | |
| 25 tokens per second of audio; context re-billed every turn; input charged while listening with proactive audio (footnote in the Gemini 2.5 section); USD | Agent Platform pricing | 26 Sep 2026 | |
| Regions us, eu, us-central1; GA; release date 24 Sep 2026 | Agent Platform Gemini 3.8 Live | 26 Sep 2026 | |
| 24 supported languages; partners Daily, LiveKit, Twilio, Voximplant | Agent Platform Live API | 26 Sep 2026 | |
| gpt-live-1 generally available 10 Sep 2026; gpt-realtime-2.1 released 6 Jul 2026 | OpenAI | API changelog | 26 Sep 2026 |
| $0.05/min billed per second; Tier 1–5 limits of 25, 50, 200, 300, 500 concurrent sessions | OpenAI | gpt-live-1 model page | 26 Sep 2026 |
| gpt-realtime-2.1 $32.00/$0.40/$64.00 audio, $4.00/$0.40/$24.00 text; mini $10.00/$0.30/$20.00 audio, $0.60/$0.06/$2.40 text; gpt-6-luna $0.10/$0.50; 10% regional uplift | OpenAI | API pricing | 26 Sep 2026 |
| 128,000-token context window for gpt-realtime-2.1 | OpenAI | gpt-realtime-2.1 model page | 26 Sep 2026 |
| 1 token per 100 ms user audio, 1 per 50 ms assistant audio; whole conversation re-sent; caching best-effort; GPT-Live bills silence | OpenAI | Voice cost guide | 26 Sep 2026 |
| SIP endpoints incl. EU; outbound SIP via Live API only; 3-minute ringing, 2-hour connected call; TLS, Opus, SDES-SRTP | OpenAI | Telephony and SIP guide | 26 Sep 2026 |
| GPT-Live G.711 μ-law and A-law at 8 kHz; one format per session | OpenAI | WebSocket guide | 26 Sep 2026 |
| Realtime formats: PCM 24 kHz only, G.711 μ-law, G.711 A-law | OpenAI | Realtime API reference | 26 Sep 2026 |
| Realtime session maximum 60 minutes | OpenAI | Realtime conversations guide | 26 Sep 2026 |
| server_vad and semantic_vad settings | OpenAI | Voice activity detection guide | 26 Sep 2026 |
| 128,000-token GPT-Live context, replacement at 90% with 8,192 tokens; 12 additional voices incl. two Australian English | OpenAI | Managing GPT-Live sessions | 26 Sep 2026 |
| “GPT-Live decides when to speak”; no end-of-response event | OpenAI | Migrate to GPT-Live | 26 Sep 2026 |
| Start backend with gpt-6-luna; client delegation accepts any agent | OpenAI | Delegation and tools | 26 Sep 2026 |
| Realtime integrations not automatically compatible with GPT-Live | OpenAI | GPT-Live partner integrations | 26 Sep 2026 |
| Live and Realtime endpoints: US and EU processing; Australia storage only | OpenAI | Your data guide | 26 Sep 2026 |
| gpt-realtime family shutdown 20 Jan 2027, notified 20 Jul 2026; gpt-realtime-mini-2025-10-06 announced 22 Apr 2026, shut down 23 Jul 2026 | OpenAI | Deprecations | 26 Sep 2026 |
| Media Streams always audio/x-mulaw, 8000 Hz, one channel | Twilio | Media Streams WebSocket messages | 26 Sep 2026 |
| Reconnect drops; turns of 100s to ~1100s; issue and PR #5497 open | Issue reporter, pipecat-ai/pipecat | #5465 | 26 Sep 2026 |
| End-call tool ~70% success, timeout warning 95%; issue and PR #7186 open | Issue reporter, livekit/agents | #5096 | 26 Sep 2026 |
| Text-only turn in audio mode; PR open | PR author, livekit/agents | PR #7390 | 26 Sep 2026 |
| OpenAI Live service for gpt-live-1 merged 10 Sep 2026 | pipecat-ai/pipecat maintainers | PR #5688 | 26 Sep 2026 |
| REFER and BYE not reaching Twilio trunk; one post, no replies | Forum poster, OpenAI Developer Community | Thread 1397428 | 26 Sep 2026 |
| Every cost in the 5-minute and 10-minute calculations | Zian AI, derived from the rates above | This page, worked calculation | 26 Sep 2026 |
Frequently asked questions
Which is better for a phone voice agent, Gemini Live API or OpenAI Realtime API?
On the telephony seam, OpenAI. Both of its voice APIs accept 8 kHz G.711 mu-law and terminate SIP, while Gemini 3.8 Live takes raw PCM over a WebSocket and returns 24 kHz PCM, so you run the transcoding bridge. On model cost, our 5-minute call ranged from $0.14 to $1.87, and GPT-Live 1 at a flat $0.05 per minute was the easiest to forecast.
Does the Gemini Live API support SIP or phone numbers directly?
The Live API documents one protocol, a stateful WebSocket connection. Google points telephony at partners such as Voximplant and Twilio. We found no SIP endpoint on the 15 Google pages we read on 26 September 2026.
Can OpenAI make outbound phone calls over SIP?
Yes, through GPT-Live, not through the Realtime API. The OpenAI Telephony and SIP guide says outbound SIP calling must be enabled for your organisation and is available through the Live API, not the Realtime API call-creation endpoint. The trunk must support TLS signalling, Opus audio and SDES-SRTP media.
How long can a single phone call run on each API?
A Realtime API session lasts at most 60 minutes. A Gemini Live connection (session management guide) lasts around 10 minutes and an audio-only session 15 minutes without compression, so longer calls need session resumption. An outbound GPT-Live SIP call is limited to 2 hours once connected.
Is Gemini Live cheaper than OpenAI Realtime per minute?
Per fresh minute of audio, yes. Gemini 3.8 Live audio (Google pricing) is $3.00 per million input tokens and $12.00 per million output tokens; gpt-realtime-2.1 audio is $32.00 and $64.00. But both re-bill the conversation every turn and only OpenAI caches it. On our 5-minute call, Gemini 3.8 Live cost $0.29, or $0.43 if line audio is billed while listening, and gpt-realtime-2.1-mini cost $0.14 when its cache held and $0.56 when it did not.
Can I keep caller audio in Australia with either API?
Not with regional processing, on the pages we read on 26 September 2026. OpenAI (Your data guide) lists the United States and Europe for its Live and Realtime endpoints, with Australia storage only. Google lists us, eu and us-central1 for gemini-3.8-live.