Two releases in the past week have quietly split the audio-and-video AI market into two different products that are easy to confuse. Google's Gemini 3.8 Live family, shipped 15 September, is built to talk — real-time speech-to-speech for voice agents. Alibaba's Qwen3.8-Omni-Flash, shipped 18 September, is built to ingest — up to an hour of continuous audio or audio-video per call inside a million-token context. Picking the wrong one is the most common and most expensive mistake teams make in this category. Here is how to tell them apart and choose deliberately.
First, Separate Live Conversation From Bulk Understanding
The distinction is latency, not capability.
Live conversation models are optimised so that a human is waiting. Gemini 3.8 Live is a native speech-to-speech model that processes visual context in near real time, switches automatically among 97 supported languages, and — the genuinely new part — executes third-party tool calls in the background while it continues speaking. Its sibling, Gemini 3.8 Live Extended Thinking, reasons and speaks simultaneously, using filler cues like "Let me check that…" and progress updates during longer jobs.
That behaviour breaks a design assumption most voice stacks were built on. The end of a spoken response no longer means the underlying task has finished. If you are building a voice agent, your state machine needs to handle "still working, still talking" as a first-class condition.
Bulk understanding models assume nobody is waiting in real time. Qwen3.8-Omni-Flash accepts a maximum input of 991,808 tokens in non-thinking mode and 983,616 in thinking mode, with up to 131,072 output tokens and a maximum reasoning length of 262K. Crucially, it handles up to one hour of continuous audio or audio-video per call — which means meeting recordings and long video no longer need chunking, and you stop losing cross-chunk context in your summaries.
Second, Do the Cost Arithmetic Before the Quality Comparison
Both releases moved price aggressively, in different units.
For live voice, Gemini 3.8 Live is billed per minute: $0.005 per minute for audio input and $0.018 per minute for output. Extended Thinking additionally charges for reasoning tokens and for extra inputs such as video and documents — budget for that separately, because reasoning volume is the least predictable line item in a voice agent.
For bulk ingestion, Qwen3.8-Omni-Flash is billed per token: international QwenCloud pricing lists $0.15 per million input tokens, $0.47 per million output tokens, and $0.016 per million cached input tokens. The more consequential change is relative: Qwen reports the API price per hour of audio input fell by more than 98%, and per hour of audio-visual input by more than 93%, against Qwen3.5-Omni-Plus.
A practical rule: if your workload is measured in concurrent conversations, price per minute. If it is measured in archives processed, price per token — and make cache-hit pricing central to your design, because it is roughly a tenfold discount.
Third, Check the Benchmarks That Match Your Job
Headline scores rarely map to the task you actually have.
- On voice, Gemini 3.8 Live Extended Thinking scored 82.6 on the Artificial Analysis Speech to Speech Quality Index, ahead of GPT-Live-1-Astra and Grok Voice Think Fast 2.0. It also posted 68.6% on tau-Voice, 35.1% on Sierra's tau-Voice-banking benchmark and 97.7% on Big Bench Audio. Standard Gemini 3.8 Live ranked second on Speech Agent Arena and first on ServiceNow's EVA-Bench.
- On omni-modal ingestion, Qwen3.8-Omni-Flash improved its average by more than 25% across 29 evaluations versus Qwen3.5-Omni-Plus, with agentic gains of 36.5 points on WildClawBench-MM and 22.3 points on AgenticVBench, and 69.6 on UniClawBench. Agentic perception lifted OmniVideoBench from 63.4 to 67.8 while using about 45.7% fewer tokens.
That last figure is the one to internalise. Token efficiency at long context matters more to your monthly bill than a two-point benchmark lead.
Why It Matters
Most teams currently run a brittle pipeline: separate speech recognition, then a text model, then separate speech synthesis. Both of these releases collapse that pipeline, and each collapses a different part of it. Consolidating removes latency, removes transcription error propagation, and removes the engineering surface that tends to rot fastest.
The trade-off is lock-in. Neither model offers open weights. Gemini 3.8 Live and Extended Thinking are hosted only. Qwen3.8-Omni-Flash launched without open weights, available through QwenCloud, Alibaba Cloud Model Studio, Qwen Studio and Qwen Chat, with developer access via DashScope using OpenAI-compatible client libraries, served from Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia. If your compliance posture requires self-hosting, neither is an option today — and Qwen's underlying Qwen3.8-Flash-Next architecture, which did ship open weights in August, is the fallback to evaluate.
A Short Selection Checklist
- Is a human waiting? Yes → live speech-to-speech. No → long-context omni-modal.
- Do tool calls run long? If so, design for speech that continues while work proceeds.
- Is your cost driver minutes or archives? Price in the matching unit, then model cache hits.
- Do you need self-hosting? If yes, filter to open-weight options first; capability comes second.
- Where must data reside? Confirm the serving region list against your obligations before you build.
One caveat worth carrying into any capacity plan: "one million tokens" does not translate to unlimited audio or video minutes. A context window covers tokenised text and signal within a single request, and the real ceiling for audio and video shifts with file size and encoding. Test with your own recordings before committing.
