Gemini 3.8 TTS vs ElevenLabs: Nearly Four Times Cheaper, Until January
The short version: on price this is not close, and it is not close in a way that is easy to miss because the two companies quote in different units. Google bills per token and converts it for you: $0.00225 per 10 seconds of audio, about $0.81 an hour of finished speech through 2026-12-31, then double. ElevenLabs' own pricing page claims low-latency TTS "as low as 5c/minute" — $3.00 an hour — and its cost calculator recommends the $11/month plan for 50 minutes of audio, roughly $0.22 a minute. Two catches before you migrate: the Google promo price doubles on January 1, and on the free tier Google says your content is used to improve its products. On quality the 146-comment thread splits cleanly: good enough for narration and read-aloud, still not convincing for character work.
The per-minute math, in one table
Google's page states its own conversion so there is nothing to reverse-engineer: audio tokens are 25 per second of speech. An hour of audio is therefore 90,000 audio tokens, and the text that generates it is a rounding error (about 12,000 tokens, well under a cent an hour).
| What you buy | Published rate (read 2026-09-25) | Per hour of audio |
|---|---|---|
| Gemini 3.8 Flash TTS | $0.50 text / $9.00 audio per 1M — $0.00225 per 10s, through 2026-12-31 | $0.81 |
| Gemini 3.8 Flash TTS, from 2027-01-01 | $1.00 text / $18.00 audio per 1M — $0.0045 per 10s | $1.62 |
| Gemini 3.8 Flash-Lite TTS | $0.50 text / $6.00 audio per 1M — $0.0015 per 10s, through 2026-12-31 | $0.54 |
| Gemini 3.8 Flash-Lite TTS, from 2027-01-01 | $1.00 text / $12.00 audio per 1M — $0.003 per 10s | $1.08 |
| ElevenLabs, Business tier claim | "Low-latency TTS as low as 5c/minute" | $3.00 |
| ElevenLabs, implied by their own calculator | Creator plan $11/month recommended for 50 minutes of generated audio | ≈$13.20 |
Which is to say: Google's Flash model is roughly 3.7× cheaper per minute than ElevenLabs' published floor, and Flash-Lite is about 5.5× cheaper. Even after the January increase, Flash-Lite at $1.08 an hour stays under ElevenLabs' Business figure. There is one honest caveat in that comparison — the 5c/minute number is ElevenLabs' low-latency enterprise claim, not what a $22 Creator subscriber pays, which is why we put both of their numbers in the table. Credits also buy more than raw audio on their side (dubbing, agents, voice changer), so per-minute price is not the whole product.
Catch one: on the free tier, you are the product
The free tier is genuinely free of charge and it is the same models. It also carries a data term, printed on the same page: "Used to improve our products: Yes" for free, "No" for paid. Paid also unlocks context caching at $0.125 per million tokens (doubling in 2027) and the batch API. If you are converting blog posts to audio for your own ears, the free tier is a good deal. If you are voicing customer-facing content or reading anyone's recordings aloud, it is the wrong tier, and that decision is worth more than the price difference.
Catch two: the price doubles on January 1, 2027
Every line in Google's Gemini 3.8 TTS row has two prices: one "through December 31, 2026" and one "starting January 1, 2027", where text, audio and caching all exactly double. The promo is long enough to look like the price and short enough to bite: fifteen months from launch, an audiobook pipeline built at $0.81 an hour runs at $1.62. That is still cheaper than the alternatives below, and it belongs in your model rather than in a footnote you find next January.
What the 146-comment thread says it is good at
Narration, mostly. The most useful comment comes from someone who has been shopping in this category for years: "I've been trying to track the SOTA for years switching between wavenet on Gcloud, to Azure, to ElevenLabs, and now Fish.audio. This is damn good" (LarsDu88, HN). And the price reaction was immediate: "Great price at least until December 31 too" (drewbitt, HN). Simon Willison built a playground within a day and reported that "the conversation mode is neat, and it's very expensive — most of my experiments have cost less than a cent" (simonw, HN) — a sentence that says the opposite of what it meant to.
Character work and accents, not yet. "Seems like voice actors are safe for now. This is technologically incredible, but the results are really not very good, and usually not particularly close to the prompt. In basically all of these examples some core part of the prompt is completely ignored" (burkaman, HN). A second reader on the same demo: "Still sounds AI, you can tell they exaggerate all the tone and trailing 'high scoring expressive sounds' like your job depends on it" (m3kw9, HN). And the accent problem is older than this model: "I've never found a TTS that does convincing British accents. They all sound like Americans putting in their best fake British accent" (andrewstuart, HN).
The rollout, meanwhile, was sloppy. "It's giving me an error when I try to generate a voice with Voice Design in AI Studio. It also says voice replication isn't available in my region... And there's no pricing listed anywhere" (nater5000, HN). Region locks and empty "use case" categories on launch day are a support-ticket tax on an otherwise cheap API.
The gaps that will cost you an afternoon
- Three Google platforms, three feature sets. "There's no alignment across the three platforms they have, consumer, prosumer, cloud... the models don't even have the same capabilities across platforms" (rcr-anti, HN). If your org licenses through Cloud, check that the model you want is the model you get.
- Consent recordings are stored somewhere. Google requires "a verbal consent recording from the voice owner that matches the reference speaker before a voice can be created" — and a reader asks the obvious follow-up: "How long is this stored? What could go wrong?" (accountrequired, HN).
- No browser extension. The most requested use case in the thread is turning articles into audio on the move (seemaze, HN), which no vendor ships as a first-class product yet.
- Conversation mode versus the API. Several readers wanted to know whether the two-speaker conversation mode is callable "like ElevenLabs offers" before re-planning anything around it (nitroedge, HN).
Who should switch, and who should stay
Switch if:
You generate bulk audio and currently pay per character: audiobook drafts, article-to-podcast, course narration, internal training material. The API is metered, the Lite model is $0.54 an hour, and 2,000+ stock voices across 100+ languages cover the boring cases. Also switch if you need regional accents or two-speaker screenplay pacing and were going to pay a human to read it.
Stay if:
Your voice work sells a character. The thread's own verdict is that prompts get partially ignored on expressive material, and the demo exaggerates in a way that reads as synthetic. Stay if you depend on ElevenLabs' dubbing, agent and studio products rather than raw TTS — Google is selling a model, not a workflow. And stay on paid tiers if your audio cannot be used to improve anyone's products.
The honest bottom line.
Google priced the marginal minute of speech at roughly a quarter of ElevenLabs' published floor and then doubled it in fifteen months, which tells you the launch price is a customer-acquisition budget rather than a cost structure. Take the discount, build so that the January number still works, and keep the character work with the vendor that has spent years on it. Our ElevenLabs alternatives breakdown covers the $5 hosted tiers if the goal is simply to stop paying $22, and the dictation guide covers the other direction — the same models listening instead of speaking.
ElevenLabs alternatives, priced → Where speech gets transcribed →
FAQ
How much does Gemini 3.8 Flash TTS cost? $0.50 per million text tokens and $9.00 per million audio tokens through December 31, 2026 — Google's own conversion is $0.00225 per 10 seconds, or about $0.81 an hour of audio. From January 1, 2027 it doubles to $1.00 text and $18.00 audio.
Is Gemini TTS cheaper than ElevenLabs? On published rates, yes: ElevenLabs' page claims low-latency TTS "as low as 5c/minute" ($3.00 an hour) and its own calculator pairs 50 minutes of audio per month with the $11 Creator plan (about $0.22 a minute), against Google's $0.81 an hour for Flash and $0.54 for Flash-Lite.
Is the free tier usable for production? Google marks free-tier content "Used to improve our products: Yes" and paid-tier content "No". Free is free; whether it is appropriate depends on whose voices and words you are sending.
Can I clone a voice with it? Yes, from a 30-second sample, with a verbal consent recording from the voice owner required, SynthID watermarking on every clip, and C2PA credentials attached.
Method: prices read from Google's Gemini API pricing page and ElevenLabs' pricing page on 2026-09-25; model claims from Google's launch post on blog.google (2,000+ voices, 100+ languages, Hume AI Voice Design Benchmark 71.4, SynthID and C2PA); user quotes verbatim from the 146-comment Hacker News thread on the release (327 points, 2026-09-24) with comment links. Per-hour figures use Google's own 25-tokens-per-second note. We take no payment from vendors.