Three real-time voice AI APIs shipped new versions inside a single September, and none of them agree on how much a minute of conversation should cost. OpenAI pushed gpt-realtime-2.1 and a cheaper gpt-realtime-2.1-mini variant onto its pricing page on September 26, 2026. Google’s Gemini 3.8 Live became the default low-latency voice model in the Gemini API on September 15. And Alibaba’s Qwen team launched Qwen3.8-Omni-Flash on September 14, publishing its own first-token and first-audio-packet latency tables down to the millisecond. For developers building voice agents, call-center automation, or speech-to-speech translation tools, the question is no longer “which chatbot is smartest” — it’s which of these three sockets you plug your product into, and what that decision costs per minute at scale.
This comparison lays out the specs, the pricing, the latency numbers each vendor and independent tester has published, and a practical migration path if you need to move between them. The gap between the cheapest and most expensive option, on a like-for-like per-million-token basis, runs as high as 213x — a number worth understanding before you commit an architecture to one provider.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What a Real-Time Voice AI API Actually Does
A real-time or “speech-to-speech” API keeps a persistent, full-duplex connection open between an application and a model so that audio can stream in and out continuously, rather than being chopped into separate speech-to-text, language-model, and text-to-speech calls. That architecture is what makes a voice agent feel like it’s actually listening rather than processing a transcript after the fact. All three APIs in this comparison — OpenAI’s Realtime API, Google’s Gemini Live API, and Qwen’s Omni-Flash realtime endpoint — solve the same core problem: a long-lived WebSocket session that accepts streaming audio, runs voice-activity detection to figure out when a person has stopped talking, generates a response, and streams synthesized audio back before the model has even finished “thinking.”
Where they diverge is in the details that matter for production systems: event schemas, audio sample rates, how interruptions (“barge-in”) are handled, how long a session can stay open, and — as the numbers below show — how the provider charges for every minute of audio that passes through the pipe.
Meet the Three Contenders
OpenAI Realtime API (gpt-realtime-2.1)
OpenAI’s Realtime API line traces back to a 2024 preview, but the current flagship is gpt-realtime-2.1, alongside a lower-cost gpt-realtime-2.1-mini. The underlying architecture jump happened with GPT-Realtime-2, announced May 7, 2026, which extended the model’s context window from 32,000 to 128,000 tokens and added specialized sibling models — GPT-Realtime-Translate for live speech translation and GPT-Realtime-Whisper for transcription-only workloads, both billed per minute rather than per token. OpenAI’s own developer documentation describes the Realtime API as enabling low-latency, multimodal interactions including speech-to-speech conversation and real-time transcription. Full rate details for every model in the family are listed on OpenAI’s public pricing page, which was last updated September 26, 2026.
OpenAI’s pricing structure rewards developers willing to trade some capability for cost: gpt-realtime-2.1-mini cuts audio-input pricing from $32 to $10 per million tokens and audio-output from $64 to $20, a straightforward way to pilot a voice agent before committing to the flagship tier. That said, neither tier competes on price with the two products below it — it’s OpenAI’s context window and existing tooling ecosystem that carry the argument for choosing it over cheaper alternatives, not raw per-minute cost.
Google Gemini Live API (gemini-3.8-live)
Google’s Live API has moved through several native-audio generations over the past year, and as of September 15, 2026, gemini-3.8-live is the default option for low-latency voice-agent and realtime-dialogue use cases, with a Gemini 3.8 Live Extended Thinking variant available for tasks that need deeper reasoning before responding. Google positions the Live API as enabling bidirectional voice and video interaction rather than audio-only exchange. If you’re setting up Gemini’s API for the first time, our Gemini 3.8 Flash API setup tutorial walks through authentication and endpoint configuration that applies to the Live variant as well.
Qwen3.8-Omni-Flash
Alibaba’s Qwen team launched Qwen3.8-Omni-Flash on September 14, 2026, an omnimodal model built to accept text, image, video, and audio input through a single realtime interface and to return text and audio output, according to Qwen’s official launch post. It ships with a 1-million-token context window and international Model Studio pricing that undercuts both US rivals by a wide margin, according to TechNode’s coverage of the launch. It’s worth noting this is a different model from the text-focused Qwen comparisons already out there — if you want to see how Omni-Flash stacks up against Qwen’s own flagship reasoning model and Claude’s latest release on coding and text benchmarks, that’s covered in our Qwen3.8-Omni-Flash vs Qwen 3.8 Max vs Claude Fable 5.1 comparison. This article focuses specifically on Omni-Flash’s realtime voice capability.
Full Specifications Compared
The table below pulls together the specifications each vendor has published for its current realtime/live voice model as of late September 2026. Where a figure comes from a third party rather than the vendor itself, that’s noted in the text following the table.
| Specification | OpenAI gpt-realtime-2.1 | Gemini 3.8 Live | Qwen3.8-Omni-Flash |
|---|---|---|---|
| Latest version release date | Pricing page updated Sep 26, 2026 (architecture from GPT-Realtime-2, May 7, 2026) | Sep 15, 2026 | Sep 14, 2026 |
| Context window | 128,000 tokens (GPT-Realtime-2 generation) | Not separately published for Live; standard Gemini 3.8 context applies per session | 1,000,000 tokens |
| Input modalities | Audio, text | Audio, video, text (bidirectional) | Text, image, video, audio |
| Output modalities | Audio, text | Native audio, text | Text, audio |
| Audio input pricing | $32 / 1M tokens ($10/1M on mini) | $3 / 1M tokens (~$0.005/min) | $0.15 / 1M tokens |
| Audio output pricing | $64 / 1M tokens ($20/1M on mini) | $12 / 1M tokens (~$0.018/min) | $0.47 / 1M tokens |
| Cached input discount | $0.40 / 1M tokens ($0.30 on mini) | Not separately published for Live audio | $0.016 / 1M tokens |
| Max session length | 60 minutes per connection (no documented resumption) | ~15 minutes default, extendable via context window compression; resumption tokens valid ~2 hours | Not separately published |
| Turn detection | semantic_vad, server_vad, or manual | automaticActivityDetection (Low/Medium/High sensitivity) | Server-side, agentic delivery framing |
| Companion specialist models | GPT-Realtime-Translate, GPT-Realtime-Whisper (per-minute billing) | Extended Thinking variant for deeper reasoning | None published separately |
| Reported language coverage | Not separately published for the realtime line | 24+ languages (Gemini TTS figure, per third-party testing) | 60 languages tested on FLEURS ASR / speech-to-speech benchmark |
| Open weights | No | No | Reported as open-weight by Qwen’s launch material |
Two things stand out immediately. First, Qwen’s 1-million-token context window matches what Anthropic and OpenAI offer on their largest text models but is unusual for a realtime voice product — most voice sessions never need anywhere near that much context, but it matters for agentic workflows that keep long tool-call histories alive during a call. For a sense of how context-window claims compare across the leading chat models generally, see our ChatGPT vs Claude vs Gemini memory capacity comparison. Second, Google’s decision to publish per-minute pricing while OpenAI and Qwen publish per-token pricing makes direct comparison harder than it should be — a problem covered in detail in the next section.
Pricing Breakdown: Where the 213x Gap Comes From
Comparing prices across these three APIs takes some care, because each vendor bills differently. OpenAI and Qwen both publish per-million-token rates for audio input and output. Google instead quotes Gemini 3.8 Live in dollars per minute, which it says is derived from an underlying $3 per million input tokens and $12 per million output tokens. That inconsistency means any single “X times cheaper” headline number should be read as directional rather than a perfect apples-to-apples conversion, since audio tokenization schemes differ by vendor — Qwen, for instance, tokenizes audio at roughly 7 tokens per second, a rate OpenAI has not published for its own audio stream. Google explains its own per-minute conversion methodology in its developer announcement for building real-time voice applications with Gemini audio.
| Provider / Model | Audio Input | Audio Output | Cached Input | Per-Minute Equivalent (approx.) |
|---|---|---|---|---|
| OpenAI gpt-realtime-2.1 | $32.00 / 1M tokens | $64.00 / 1M tokens | $0.40 / 1M tokens | ~$0.060 input / $0.240 output (third-party estimate) |
| OpenAI gpt-realtime-2.1-mini | $10.00 / 1M tokens | $20.00 / 1M tokens | $0.30 / 1M tokens | Not separately benchmarked |
| OpenAI GPT-Realtime-Translate | $0.034 / minute (flat rate) | N/A | $0.034 / minute | |
| OpenAI GPT-Realtime-Whisper | $0.017 / minute (flat rate) | N/A | $0.017 / minute | |
| Gemini 3.8 Live | $3.00 / 1M tokens | $12.00 / 1M tokens | Not published | $0.005 input / $0.018 output |
| Qwen3.8-Omni-Flash | $0.15 / 1M tokens | $0.47 / 1M tokens | $0.016 / 1M tokens | ~$0.0038 / hour input at 7 tokens/sec (derived, not vendor-quoted) |
On raw published input-token pricing, OpenAI’s $32-per-million-token rate for gpt-realtime-2.1 sits 213 times higher than Qwen3.8-Omni-Flash’s $0.15-per-million-token rate. On output tokens, the gap is smaller but still stark: $64 versus $0.47, a 136x difference. Those ratios compare official list prices for audio-input and audio-output tokens as published by each company, though — again — the two vendors count “tokens” differently, so treat the multiple as a sticker-price comparison rather than a guaranteed real-world cost ratio.
Independent benchmarking from AutoInterviewAI, published September 16, 2026, converted OpenAI and Google’s rates to a common per-minute basis and found Gemini 3.8 Live’s blended cost (audio in plus audio out) at roughly $0.023 per minute against OpenAI’s roughly $0.300 per minute — about a 13x gap between those two specifically. At a volume of 500,000 call minutes a month, that difference works out to roughly $831,000 in annual savings by choosing Gemini over OpenAI’s flagship tier, according to the same report. If your architecture needs to route requests across multiple providers to capture savings like that dynamically, our guide to building a multi-model AI router with fallback covers the load-balancing pattern, though that guide focuses on text models rather than realtime audio streams specifically.
None of the three vendors currently publishes a confirmed free tier for its realtime voice endpoint specifically, based on each provider’s current billing documentation. Budget accordingly from the first API call.
Latency and Real-Time Performance
Latency is the metric that determines whether a voice agent feels natural or feels like talking to a walkie-talkie. Qwen has published the most granular first-party numbers of the three. Its launch material reports a time-to-first-text-token of 591.26ms for a 6-second audio input, rising slightly to 604.80ms at 12 seconds and 617.98ms at 20 seconds. Time to the first actual audio packet is higher — 978.36ms at 6 seconds, 982.74ms at 12 seconds, and 1,026.39ms at 20 seconds. For combined audio-visual input, Qwen reports 837.96ms to first text token and 1,214.73ms to first audio packet on a 6-second clip, and an audio-generation real-time factor of about 0.154, meaning the model generates audio roughly 6.5 times faster than real-time playback speed.
OpenAI and Google don’t publish directly comparable first-party latency figures for their realtime models in the same format, which is why third-party testing matters here. MindStudio’s comparison of GPT Realtime 2 against Gemini’s audio stack found GPT Realtime 2’s first audio chunk typically arriving in the 300–500ms range, while describing Gemini’s latency as competitive without giving Google’s own precise figure. AutoInterviewAI’s September 16 benchmark went further, reporting a barge-in cutoff latency of under 40ms for Gemini 3.8 Live versus under 50ms for OpenAI’s model, and a time-to-first-audio window of 180–220ms for Gemini against 200–250ms for OpenAI in its specific test harness.
It’s worth treating all three of these latency claims with some skepticism about direct comparability. Zylos Research’s July 2026 technical teardown of both protocols makes the point explicitly: reported time-to-first-audio figures for the same underlying model can range from 0.6 to more than 3 seconds depending entirely on where in the pipeline the clock starts and stops, and whether a relay server sits in the network path. A benchmark run against a cloud endpoint through a media relay will look slower than the identical model queried directly, regardless of which company built it.
Benchmark Results From Three Independent Sources
No single leaderboard — Artificial Analysis, LMSYS, or otherwise — currently ranks all three of these specific realtime models side by side using a shared, independently-run methodology. That’s a genuine gap in the public benchmark landscape as of September 2026. What exists instead is a mix of vendor-published numbers and independent testing, summarized below.
| Benchmark | Gemini 3.8 Live | OpenAI GPT-Realtime-2 | Source |
|---|---|---|---|
| Artificial Analysis S2S Index | 82.6 | 79.4 | AutoInterviewAI, Sep 16, 2026 |
| τ-Voice Agentic Task Completion | 68.6% | 52.3% | AutoInterviewAI, Sep 16, 2026 |
| Sierra τ-Voice-Banking | 35.1% | 24.8% | AutoInterviewAI, Sep 16, 2026 |
| Big Bench Audio Comprehension | 97.7% | 91.2% | AutoInterviewAI, Sep 16, 2026 |
| Voice count (native voices) | 30+ (Gemini TTS) | 6 | MindStudio comparison |
| Language coverage | 24+ (Gemini TTS) | Not separately published | MindStudio comparison |
Qwen doesn’t yet appear in this particular third-party benchmark set, since the comparison predates Omni-Flash’s September 14 launch by a matter of days. Its own reported numbers — the FLEURS ASR and speech-to-speech translation evaluation covering 60 languages including Mandarin, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Korean, and Hindi — are self-reported rather than independently verified, which is standard for a same-week product launch but worth flagging before treating any single number as gospel.
Reading across all three sources, the pattern that emerges is consistent even without a unified leaderboard: Gemini 3.8 Live’s own published pricing and the third-party benchmark data agree that it currently undercuts OpenAI on cost by a wide margin while scoring competitively or better on the agentic and comprehension tests AutoInterviewAI ran. Qwen’s figures suggest strong raw latency for a model priced far below either US competitor, though the absence of independent verification means those numbers deserve a follow-up test once third parties get access.
Language and Multimodal Support
Qwen3.8-Omni-Flash is the most explicitly multimodal of the three, built to accept text, image, video, and audio as input through one endpoint and apply the same international per-token price across all of them, according to pricing analysis from Digital Applied. That single-price-for-every-modality approach is unusual in a market where most vendors charge separately by media type. Its 60-language FLEURS evaluation set spans Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, and Urdu, among others — a genuinely global spread for a model that launched two weeks before this comparison was written.
Google’s Gemini Live API supports bidirectional audio, video, and text, per its own documentation. The Live API enables low-latency, real-time voice and vision interactions with Gemini, according to Google’s Live API documentation. That video-input capability is a meaningful differentiator for use cases like a support agent that needs to see a screen-share or a photo alongside a spoken question — something neither OpenAI’s realtime line nor Qwen’s Omni-Flash documentation currently emphasizes to the same degree for the realtime endpoint specifically.
OpenAI’s realtime family, by contrast, is more narrowly focused on audio and text, with the heavy multimodal lifting (image and video understanding) handled by its separate non-realtime model lines. The specialized GPT-Realtime-Translate and GPT-Realtime-Whisper models extend the family into transcription and translation specifically, priced and billed independently from the full-duplex conversational model.
Architecture and Protocol Differences
Underneath the marketing, these are three different wire protocols, and the differences are exactly the kind of thing that breaks a production integration if you don’t account for them up front. Both OpenAI and Gemini use WebSocket transport, but everything downstream of the connection differs. OpenAI initializes a session with a session.update client event; Gemini expects a setup message as the very first frame. OpenAI’s input audio typically runs at a configurable rate around 24kHz; Gemini’s Live API fixes input audio at 16kHz and output at 24kHz, which means any audio captured at a different sample rate needs band-limited resampling — naive rate conversion introduces aliasing that audibly degrades consonants, according to Zylos Research’s technical teardown of the two protocols.
Turn detection also diverges. OpenAI offers semantic_vad, server_vad, or fully manual turn control. Gemini uses automaticActivityDetection with Low, Medium, or High sensitivity settings, and Google’s own guidance recommends at least 500–800ms of silence before considering a turn complete — a threshold that can clip languages with naturally shorter pauses between phrases. Session length is another practical difference: OpenAI caps a single Realtime connection at 60 minutes with no documented resumption mechanism, while Gemini’s default session runs about 15 minutes but supports resumption tokens valid for roughly two hours and a context-window-compression feature to extend effective session length beyond that default.
// OpenAI Realtime API session init
{
"type": "session.update",
"session": {
"modalities": ["audio", "text"],
"voice": "alloy",
"turn_detection": { "type": "server_vad" }
}
}
// Gemini Live API setup message (first frame)
{
"setup": {
"model": "models/gemini-3.8-live",
"generationConfig": { "responseModalities": ["AUDIO"] },
"realtimeInputConfig": {
"automaticActivityDetection": { "startOfSpeechSensitivity": "MEDIUM" }
}
}
}Tool-calling flow is asymmetric too: OpenAI enforces sequential blocking, where the model emits a function_call, the client executes it and returns function_call_output, and only then does the client trigger response.create to resume generation. Gemini permits asynchronous tool use, letting generation continue while a tool call executes in the background — faster in principle, but it introduces its own class of bugs, including reports of Gemini’s native-audio models sending a turnComplete signal mid-generation without an accompanying interrupted flag, which can cut audio off mid-sentence if a client trusts the protocol signal at face value.
Migration Guide: Moving Between Realtime Voice APIs
Switching a production voice agent from one of these APIs to another is not a drop-in change — it’s closer to a protocol translation project. Here’s a practical sequence for teams evaluating a move, based on the adapter pattern documented by Zylos Research’s July 2026 protocol teardown.
- Pick a target event schema. Most migration guides treat OpenAI’s event naming as the de facto standard and translate the other provider’s events into that shape, since more existing tooling assumes OpenAI-style events.
- Map session initialization. Translate OpenAI’s
session.updateinto Gemini’ssetupmessage (or vice versa), matching modality configuration and voice selection fields explicitly rather than assuming defaults line up. - Handle audio resampling correctly. If moving from OpenAI’s ~24kHz input to Gemini’s fixed 16kHz requirement, apply a band-limited low-pass filter before decimation — a staged 3:2 ratio conversion outperforms a single large-ratio resample and avoids the aliasing artifacts that degrade consonant clarity.
- Rebuild turn-interruption logic. OpenAI requires explicit client-side truncation via
conversation.item.truncatewith precise audio timing; Gemini uses server-declarativeinterruptedflags and expects the client to manage its own playback buffer. These are not equivalent and need separate code paths. - Re-tune voice-activity detection per language. Neither protocol ships pre-tuned for every language. Validate turn-completion thresholds against real conversational audio in your target languages rather than synthetic test clips, especially for tonal languages where natural pauses run shorter than the 500–800ms Gemini recommends by default.
- Decouple tool-call state from turn state. Because Gemini allows asynchronous tool execution while OpenAI blocks sequentially, an adapter has to track tool-call completion independently of conversational turn completion to avoid race conditions.
- Add defensive handling for mid-sentence cutoffs. Build heuristics that flag suspiciously short responses immediately following a tool call, and re-prompt rather than trusting a
turnCompletesignal without a correspondinginterruptedflag. - Decide between a purpose-built adapter and a framework. A hand-rolled adapter makes sense when adding exactly one additional provider to an existing single-provider system. Once you’re supporting three or more providers, or need telephony and room infrastructure alongside the voice model itself, a framework like Pipecat (transport-agnostic), LiveKit Agents (provider-plugin model with native Gemini video support), or TEN Framework (graph-based, pluggable turn detection) becomes the more defensible investment.
- Re-run your latency and cost benchmarks on your own traffic. Published latency figures vary by 5x or more depending on measurement point and network path, so a vendor’s or reviewer’s number is a starting estimate, not a guarantee for your specific deployment region and call patterns.
Real-World Use Cases for Each API
These three APIs target overlapping but distinct deployment patterns based on their published strengths and the benchmark suites vendors and reviewers have chosen to test them against.
- Banking and financial voice agents: Sierra’s τ-Voice-Banking benchmark specifically simulates account-servicing and financial task completion over voice, and Gemini 3.8 Live’s 35.1% completion rate against OpenAI’s 24.8% on that suite (per AutoInterviewAI’s testing) suggests Gemini currently has an edge on structured, high-stakes financial dialogue tasks.
- High-volume customer support call centers: At call volumes in the hundreds of thousands of minutes a month, the roughly 13x per-minute cost gap between Gemini 3.8 Live and OpenAI’s flagship tier translates into six-figure annual savings, making cost-sensitive, high-volume support deployments a natural fit for Gemini’s pricing structure specifically.
- Compliance-heavy or long-session enterprise workflows: OpenAI’s larger 128,000-token context window on GPT-Realtime-2 and its more mature tool-calling documentation make it a safer first choice for regulated industries that need longer audit trails within a single session, even at a higher per-minute cost.
- Multilingual and cross-border deployments: Qwen3.8-Omni-Flash’s 60-language FLEURS evaluation and unified per-token pricing across text, image, video, and audio make it worth testing for products serving audiences across Southeast Asia, the Middle East, and Latin America where language coverage matters as much as raw latency.
- Screen-aware support and visual troubleshooting: Gemini Live’s native video-input support allows a spoken support conversation to include a live camera or screen-share feed, a capability neither OpenAI’s realtime line nor Qwen’s documentation currently emphasizes to the same degree for step-by-step visual troubleshooting.
- Live speech translation and transcription-only pipelines: OpenAI’s GPT-Realtime-Translate ($0.034/minute) and GPT-Realtime-Whisper ($0.017/minute) exist specifically for teams that don’t need a full conversational agent, just accurate translation or transcription at a flat, predictable per-minute rate.
Working through a concrete scale example makes the cost differences easier to internalize. A mid-sized customer support operation running 50 concurrent voice agents for eight hours a day, five days a week, generates roughly 480,000 call minutes a month. At OpenAI’s third-party-estimated blended rate of about $0.30 per minute, that workload costs roughly $144,000 a month. At Gemini 3.8 Live’s blended rate of about $0.023 per minute, the same volume costs closer to $11,000 a month — a gap wide enough to fund a small engineering team purely from the switching decision, assuming the underlying task-completion quality holds up in production testing rather than just in AutoInterviewAI’s benchmark harness. Qwen3.8-Omni-Flash’s derived hourly input cost of roughly $0.0038 suggests an even lower theoretical floor, though its output-token pricing of $0.47 per million and the absence of a directly comparable per-minute blended figure make it harder to model with the same precision until more third-party cost audits exist.
Pros and Cons
OpenAI Realtime API (gpt-realtime-2.1)
- Pros: 128,000-token context window, mature documentation and tooling ecosystem, dedicated per-minute-billed translation and transcription models, well-understood by most existing voice-AI frameworks.
- Cons: Most expensive audio pricing of the three by a wide margin, only 60-minute sessions with no documented resumption, fewer native voices (6) than Gemini’s TTS lineup (30+), no confirmed free tier.
Gemini 3.8 Live
- Pros: Roughly 13x cheaper per minute than OpenAI’s flagship tier per third-party testing, native video input alongside audio, session resumption tokens valid ~2 hours, stronger scores on the agentic-task and audio-comprehension benchmarks AutoInterviewAI ran.
- Cons: Shorter 15-minute default session length before resumption is needed, less granular first-party latency disclosure than Qwen, VAD tuning defaults (500–800ms silence) can clip tonal-language turns, reported mid-generation
turnCompletebugs requiring defensive client code.
Qwen3.8-Omni-Flash
- Pros: Dramatically lower published per-token pricing than either US competitor, 1-million-token context window, single unified price across text/image/video/audio, 60-language FLEURS coverage, reported open-weight availability.
- Cons: No independent third-party benchmark yet places it alongside OpenAI and Gemini on a shared leaderboard, session-length limits not separately published, newest of the three with the least production track record as of this writing.
What the Companies Say
OpenAI’s own developer documentation describes its offering plainly: “The OpenAI Realtime API enables low-latency, multimodal interactions including speech-to-speech conversational experiences and real-time transcription,” according to OpenAI’s Realtime API guide. Google frames its competing product in similar terms but leads with bidirectionality: “The Gemini Live API enables low-latency bidirectional voice and video interactions with Gemini,” per Google Cloud’s Gemini Live API reference documentation. Google’s developer-facing docs describe the same product slightly differently, emphasizing the vision angle specifically: “The Live API enables low-latency, real-time voice and vision interactions with Gemini,” according to Google’s Gemini API Live API overview.
Neither company’s official language makes a direct performance claim against the other — the competitive framing in this space is currently being driven almost entirely by independent testers and pricing analysts rather than head-to-head statements from OpenAI, Google, or Alibaba themselves.
How Third-Party Reviewers Are Framing the Decision
Beyond the benchmark numbers, a growing crop of comparison sites has already started publishing quick-take verdicts on this exact matchup. JustThink AI’s side-by-side comparison frames the decision in risk terms rather than pure performance: it recommends OpenAI’s Realtime API for teams that want the lower-risk, more established default, and suggests Gemini Live is worth testing specifically when a product is already built around Google-native, multimodal experiences rather than as a blanket replacement. That framing lines up with the technical picture in this article — Gemini’s cost and video-input advantages matter most when they map directly onto a use case that needs them, not as a universal upgrade.
Community comparison trackers on sites like SourceForge and Slashdot have also started listing all three APIs side by side for feature checklists, though those listings tend to reproduce vendor-published specs rather than run independent tests. For teams that need to evaluate several providers’ pricing models simultaneously rather than committing to one, the same fallback-routing logic used for text models — described in our breakdown of Grok 4.7’s API pricing relative to its rivals — applies conceptually here too: build the ability to route to a cheaper or faster provider before you need it, rather than after a price hike or outage forces an emergency migration.
Which API Should You Choose?
- Choose OpenAI’s Realtime API if your product already runs on OpenAI’s ecosystem, needs the largest available context window for long agentic sessions, or requires the specialized flat-rate translation and transcription models.
- Choose Gemini 3.8 Live if per-minute cost at scale is the deciding factor, if your use case benefits from native video input alongside voice, or if your workload maps closely to the agentic and comprehension tasks where Gemini scored higher in third-party testing.
- Choose Qwen3.8-Omni-Flash if you’re serving a genuinely multilingual audience across 60+ languages, need a single unified price across text, image, video, and audio modalities, or want to evaluate an open-weight option before committing to a closed API.
- Choose OpenAI’s GPT-Realtime-Translate or GPT-Realtime-Whisper specifically if you don’t need a conversational agent at all — just accurate, flat-rate live translation or transcription.
- Build a provider-agnostic adapter layer from day one if you expect to support more than one of these APIs, since the protocol differences documented above (audio sample rates, turn detection, tool-call flow) make single-provider assumptions expensive to unwind later.
The Verdict
There’s no single winner across every dimension here, and the data makes that clear rather than obscuring it. On raw published pricing, Qwen3.8-Omni-Flash is in a different league entirely — 213x cheaper than OpenAI on audio input tokens and roughly triple the context window of GPT-Realtime-2, though it’s also the least independently verified of the three just two weeks after launch. On a common per-minute basis measured by AutoInterviewAI, Gemini 3.8 Live undercuts OpenAI by roughly 13x while also posting higher scores on agentic task completion (68.6% versus 52.3%) and audio comprehension (97.7% versus 91.2%) in that same test. OpenAI’s gpt-realtime-2.1 remains the most expensive of the three on every pricing metric published, but it carries the largest context window among the two US models and the deepest existing tooling ecosystem, which counts for something in production environments that can’t afford to be first movers on an unproven stack.
For teams building net-new in September 2026: start with Gemini 3.8 Live if cost and video input matter most, evaluate Qwen3.8-Omni-Flash in parallel if your audience is multilingual or you want to keep an open-weight option on the table, and reserve OpenAI’s Realtime API for workloads that specifically need its longer session context or its dedicated translation and transcription models. Whichever you pick, build the protocol adapter layer described in the migration section above before you need it — not after a pricing change or an outage forces the question.
Frequently Asked Questions
Is Gemini 3.8 Live actually cheaper than OpenAI’s Realtime API?
Yes, on every published pricing metric. Google quotes Gemini 3.8 Live at roughly $0.005 per minute for audio input and $0.018 per minute for audio output, versus OpenAI’s roughly $0.06 and $0.24 per-minute equivalents on gpt-realtime-2.1, according to third-party conversion by AutoInterviewAI. That works out to a blended cost difference of roughly 13x in Gemini’s favor at the same call volume.
Which of the three has the lowest latency?
No single independently verified leaderboard ranks all three together. Qwen publishes the most granular first-party figures — 591–618ms to first text token and 978ms–1.03 seconds to first audio packet depending on input length. Third-party testing from AutoInterviewAI found Gemini 3.8 Live’s time-to-first-audio (180–220ms) edging out OpenAI’s (200–250ms) in that specific test, but Zylos Research’s protocol analysis warns that measurement point and network path can shift reported latency by several times regardless of the underlying model.
Is Qwen3.8-Omni-Flash open source?
Qwen’s launch material and subsequent coverage describe Omni-Flash as an open-weight model, which is notable for a realtime multimodal system at this capability level. Neither OpenAI’s Realtime API nor Google’s Gemini Live API is open-weight; both are closed, API-only products.
Can I switch between these APIs without rewriting my application?
Not without an adapter layer. The three APIs use different session-initialization messages, different fixed or configurable audio sample rates, different turn-detection mechanisms, and different tool-calling flows. The migration guide section above outlines the specific mapping work required, and frameworks like Pipecat, LiveKit Agents, or TEN Framework can absorb some of that translation work if you plan to support more than one provider long-term.
Do any of these APIs offer a free tier?
Based on each provider’s current billing documentation as of late September 2026, none of the three confirms a free usage tier specifically for its realtime voice endpoint. Budget for paid usage from your first test call.
Which API supports the most languages?
Qwen3.8-Omni-Flash reports the broadest tested coverage, with a 60-language FLEURS ASR and speech-to-speech translation evaluation. Gemini’s TTS lineup is documented at 24+ languages in third-party testing. OpenAI has not separately published a language-coverage figure specifically for its realtime line.
Why does OpenAI limit sessions to 60 minutes?
OpenAI has not published a stated technical reason for the 60-minute cap, and it currently offers no documented session-resumption protocol comparable to Gemini’s resumption tokens. In practice, this means long-running voice applications built on OpenAI’s Realtime API need to plan for session handoff logic at the 60-minute mark, reconnecting and re-establishing context rather than relying on a single unbroken connection.
Which model scores best on agentic voice tasks?
In AutoInterviewAI’s September 16, 2026 benchmark, Gemini 3.8 Live scored 68.6% on the τ-Voice Agentic Task Completion benchmark versus 52.3% for OpenAI’s GPT-Realtime-2, and 35.1% versus 24.8% on Sierra’s more specialized τ-Voice-Banking suite. Qwen3.8-Omni-Flash was not included in that particular test round, which predates its September 14 launch by a few days.
