GPT-Live-1 与 Gemini 3.8 TTS：一个能对话，一个只会念稿——而两者中只有一个便宜

Set G​PT-Live-1 and Gemini 3.8 Flash TTS next to each other on a per-minute meter and the answer looks obvious. G​oogle's narrator bills about 1.35 cents per minute of generated speech at the 2026 rate. O​penAI's conversational model bills 5 cents per minute of open line — and keeps billing while the backend thinks, while the user pauses, and while nothing is being said at all. Roughly a twenty-seven-fold gap, before anyone has asked whether the two are doing the same job.

They are not, and in a way that is more interesting than the usual "different categories" hand-wave. Gemini 3.8 Flash TTS cannot hear. G​PT-Live-1 cannot be batched. And the naming makes this worse than it needs to be, because the phrase G​emini 3.8 TTS — the label half the internet is using — does not correspond to any model you can call.

首先，是名称问题，因为它决定你报出的价格

There is no model called G​emini 3.8 TTS. G​oogle's own documentation uses the phrase as umbrella language for two callable endpoints, and they are priced differently enough that quoting the wrong one is a real error:

• gemini-3.8-flash-tts — 完整版本，支持 130 多种语言，文本输入每百万 token 收费 0.50 美元，音频输出每百万 token 收费 9.00 美元，2027 年 1 月 1 日起将涨至 1.00 美元和 18.00 美元。

• gemini-3.8-flash-lite-tts — 支持 101 种语言，输入价格不变，每百万音频输出 token 收费 6.00 美元，之后涨至 12.00 美元。

• 两者均提供 Batch 和 Flex 费率，为 Standard 价格的一半（输入 $0.25，输出 $4.50 或 $3.00），以及 Priority 费率，为 1.8 倍（输入 $0.90，输出 $16.20 或 $10.80）。

音频按每秒 25 个 token 计费，每分钟的数字正源于此：25 个 token × 60 秒 = 每分钟 1,500 个音频 token，因此每百万 $9.00 折合每分钟约 $0.0135。Flash-Lite 约为 $0.009。

The status of the model is genuinely unsettled and worth saying out loud, because G​oogle's own surfaces disagree. The G​emini API changelog calls both models generally available. The API model list labels them "New Stable." G​oogle Cloud's text-to-speech documentation describes them as available in Preview through the G​emini Enterprise API only. The announcement blog post, published on 23 September 2026, uses neither word. Anyone writing a procurement note this month should treat the GA label as the changelog's claim rather than a settled fact.

当你调用每个模型时，它实际上会做什么

Gemini 3.8 Flash TTS 接收文本并返回音频。它没有音频输入，没有需要接的话轮，没有被中断的概念，也没有可交回的转录文本，因为它从未听到过任何东西。一元端点返回 24 kHz 单声道 WAV；流式端点返回无头部的 16 位 PCM，接受 mulaw 和 alaw，且采样率可配置。内联舞台指示以尖括号事件的形式写入脚本——<laugh>、<sigh>、<breath>、<short pause>——除此之外，输入文本被视为逐字转录稿，而非指令。Live API 完全不支持这些模型。

G​PT-Live-1 does the opposite job. It is full duplex: it processes incoming audio while producing outgoing audio, decides several times a second whether the user has finished talking, and returns speech-recognition transcripts as part of the session. Its inputs are audio and text; its outputs are audio and text; images and video are unsupported. It streams and supports function calling but not structured outputs, and knowledge cut-off is 31 July 2025. Only one endpoint serves it — v1/live/sessions — and Chat Completions, Responses, Realtime and Batch are all out.

That last detail is the one that ends the comparison for a lot of workloads. If your product renders a script, G​PT-Live-1 is neither the cheap option nor the correct one. If your product answers a phone, Gemini 3.8 Flash TTS is one third of a pipeline that still needs an ASR model and a language model bolted on either side — and the whole point of G​PT-Live-1 is that the seam between those three parts is where conversations break.

这些数字，以及它们属于谁

On Artificial Analysis's Provider Voice Arena — a blind-vote Elo board over each vendor's own native voices — G​oogle's new models land mid-field rather than at the top:

• Gemini 3.8 Flash TTS — 在 2,783 个样本上，以 1,275 的 Elo 分数位列第 5 名，95% 置信区间约为 1,260 至 1,290，标价为每百万字符 $16.5。

• Gemini 3.8 Flash-Lite TTS —— Elo 1,248，位列第 8，区间约为 1,234 至 1,262，每百万字符 11.0 美元。

• Above both — Eleven v4 Turbo at 1,330, Eleven v4 at 1,329, Q​wen-Audio-3.1-TTS-Plus at 1,296 and Cartesia Sonic 3.6 at 1,285.

G​oogle's own claim is stronger than that board, and it is a vendor claim: the 23 September post says the models took first place on Hume AI's Voice Design Benchmark at 71.4 and led Hume's accent modelling at 60.8, plus top blind-preference positions in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. Those are G​oogle-reported placements on someone else's evaluation, not a reproduction.

Flipping to the conversational side, G​PT-Live-1's numbers come from a different board entirely. Artificial Analysis's speech-to-speech index — a weighted average of speech reasoning, agentic performance, arena preference and task success — now lists G​PT-Live-1 under named backend configurations rather than as one row: G​PT-Live-1 (Astra, medium) and G​PT-Live-1 (Sol, low) sit at the top of that index, with G​rok Voice 3.0 leading the field. That is a useful reminder that in this architecture the voice model is only half the row; the delegated backend moves the score.

One number that matters and does not exist anywhere: G​oogle publishes no official latency figure for either TTS model. Every time-to-first-audio number in circulation is third-party — one aggregator's p50 sits around 3.47 seconds for Flash TTS and 2.66 for Flash-Lite — and none of it is directly comparable to G​PT-Live-1's conversational cadence, because they measure different things.

语音设计既是二者交汇之处，也是它们再度分道扬镳之处。

两家厂商现在都能让你用文字描述一种声音。这是2026年才有的新功能，也是这两款产品最接近彼此重叠的地方。

G​oogle's path runs through the G​emini API: a prompt-based voice persona saved persistently as a voice_ ID, voice replication from about a 30-second sample with a consent check, SynthID watermarking and C2PA credentials attached, and a cap of 200 stateful custom voices per project with one-year retention from last use. The written description covers role, accent and vocal traits.

O​penAI's path, released on 16 October 2026 to all API customers, is a POST to /v1/audio/voices with type: "prompt", a name, the description and "model": "auto". Creation is free; using the voice bills at normal session rates. It returns metadata only — no audio, no preview. The qualities you can steer are accent, age, presentation, pitch, timbre, texture, loudness, speech rate and distinctiveness, all expressed as ordinary language with no numeric parameters.

两个差异决定了你实际可以使用哪一个：

• Prompt-generated voices on G​PT-Live work only with G​PT-Live. They are explicitly incompatible with the Text to Speech, Realtime and Chat Completions APIs. G​oogle's custom voices are, by contrast, voices for a rendering engine, which is what makes them reusable across a catalogue.

• The organisation cap on O​penAI's side is 100 prompt-generated voices, and it applies to voices generated on behalf of your customers too. G​oogle's ceiling is 200 per project. If you are a platform minting a voice per customer, both numbers are your real product limit, not a footnote.

There is also a difference in what you get back. A G​oogle voice arrives with SynthID and C2PA provenance attached. An O​penAI prompt-generated voice arrives as an ID, a creation timestamp and your own prompt text — and the documentation says outright that generated voices are not guaranteed to be unique. Keep the description under version control; it is the only record of what you asked for.

当价格差距不再是 27 倍时

The per-minute comparison is the one that gets quoted and it is the one that misleads, because G​PT-Live-1's meter is not measuring the same resource. Three things move the real ratio:

• Idle time is billable. A voice session is billed per second for as long as it is open, including while the backend model is working and while the user is silent. G​oogle's meter only advances when audio is generated. A support call where the user talks for four minutes and your backend thinks for two is six billable minutes on Live and fewer on the renderer — before you have counted the ASR and language-model legs the renderer still needs.

• 输出长度只在其中一方是可控的。渲染任务是已知量：你可以在生成之前就为一篇 900 字的文章定价，而批量费率还能在此基础上再减半。对话则不然。

• The backend is a separate bill. Delegated reasoning and tool use on G​PT-Live-1 are charged at normal model rates on top of the $0.05 a minute, and O​penAI's own cost guidance frames the lever as making the user finish sooner — for example, running a faster backend with parallel tool calls so the conversation ends a minute earlier, saving $0.05. The voice rate is fixed; the session length is the dial.

That last point is where a routing layer stops being a nice-to-have. Client delegation lets you point a G​PT-Live session at your own agent or service, which means the delegated backend can be any model — and the useful version of "any model" is one endpoint and one key rather than a contract per provider. OrcaRouter fronts 200-plus models at provider list price with no markup, so a vendor price cut on the backend side reaches you the day it lands, and automatic failover covers a provider outage without the caller hearing it. If you are building the delegated path, the backend choice is where your cost actually lives; the $0.05 a minute is a constant.

To be precise about the boundary: OrcaRouter does not host G​PT-Live-1 or the G​emini 3.8 TTS models. What it routes is the reasoning side of a delegated Live session.

该选哪一个

If you are rendering — articles, documentation, video narration, IVR prompts written in advance — Gemini 3.8 Flash TTS is the correct tool and G​PT-Live-1 is roughly 27 times too expensive per audio minute and structurally incapable of the job in batch. Take Flash-Lite if 101 languages cover you, take Batch or Flex rates if latency is not the constraint, and model the January 2027 price doubling from the start, because G​oogle has published it in the same table rather than burying it.

如果你是在进行一场对话，那么 TTS 对比就是一种范畴错误，而该诚实提出的问题是另一个：被委托的后端能否足够快地给出答案，让会话保持简短。这个数字才决定你的账单，而它正是没人会放上发布会幻灯片的那一个数字。

And if you are simply trying to work out what "G​emini 3.8 TTS" costs — the answer is $9.00 per million audio tokens, on one of the two models that name refers to, until the end of 2026.

这一切都并非假设：像本次这样的发布背后那些有据可查的 AI 故障与智能体事件，均收录于 OrcaRouter AI 事件档案库中，其中每条记录都附有自己的来源、严重程度、置信度评级和存疑标记。