Gemini 3.8 TTS vs Breeze TTS 2: The Licence, Not the Leaderboard, Decides This One

Fifty-six Elo points separate Gemini 3.8 Flash TTS and Breeze TTS 2 on the Artificial Analysis Provider Voice Arena, and they are not inside each other's error bars: The vendor's model sits at rank 2 with an Elo of 1,260 across 1,999 samples and 8 arena voices, released September 23, 2026; Breeze TTS 2 sits at rank 9 with 1,204 across 1,390 samples, released August 2026. On price the ordering flips hard — Breeze is listed at $34.00 per million characters normalised against $16.50 for Flash TTS, so the model that wins on the board loses on the bill by roughly double.

If that were the whole comparison, it would be a short piece and you would pick based on how much you trust a 56-point gap. But the two models are not the same kind of object. Breeze TTS 2 is a 3-billion-parameter open-weights model you can download, inspect and run yourself, with a licence that turns out to matter more than its position on a leaderboard. Gemini 3.8 Flash TTS is an API endpoint with voice design and two-speaker dialogue and no weights at all. Deciding between them is a licensing and deployment question wearing a benchmark's clothes.

What each one actually is

Get the shape of each model straight before comparing numbers, because they differ on almost every axis that is not Elo.

• Distribution — Breeze TTS 2 is downloadable: weights published on Hugging Face as BreezeBlue/Breeze-TTS-2 alongside PyTorch code. Gemini 3.8 Flash TTS is API-only, reachable through the Gemini API and the Google surfaces named in the September 23 announcement.

• Size — Breeze TTS 2 is a 3B-parameter model, which is the reason it can be self-hosted at all. Google has not published a parameter count for either Gemini 3.8 TTS model.

• Voice creation — Gemini 3.8 Flash TTS does voice design from a written description and voice replication from a 30-second sample. Breeze TTS 2 ships a voice library the vendor counts in the thousands and its own voice-design path; the vendor-reported role-fit and direction scores for that path are BreezeBlue's own and have no independent counterpart.

• Dialogue — Gemini 3.8 Flash TTS generates two-speaker dialogue in a single call, with turn-level styling. Breeze TTS 2 is a single-voice model; a conversation is two runs stitched.

• Languages — Breeze TTS 2 covers 50 languages by the vendor's count. Google's announcement counts 130 for the Flash model, but the two counts are not measuring the same thing and should not be read as 130-versus-50.

• Latency — BreezeBlue claims sub-40 ms streaming latency; Google publishes no equivalent figure for Flash TTS. Neither number is independently verified, and only one of them exists.

The licence is the real fork in the road

This is the part the leaderboard cannot show you, and it is the reason the two models are not interchangeable even where their quality overlaps.

Breeze TTS 2's weights ship under the BreezeBlue research and non-commercial licence. The source code in the same repository is Apache 2.0. Those two facts sit next to each other on the model card and get collapsed into "open source" in most coverage, which is wrong in the direction that costs money: the code being Apache 2.0 does not make the weights Apache 2.0, and commercial use of the model runs through the hosted API rather than through a self-hosted checkpoint.

Google's side has no licence question at all, because there is nothing to license. You are buying inference. What you give up is the thing open weights are usually bought for — the ability to run the model where the data is, to pin a version and keep it, and to stop paying per character.

So the honest framing is not open versus closed. It is:

• If your deployment has to keep audio inside your own infrastructure, Breeze TTS 2 is the only one of the two that offers a path, and the path is a research licence. Check with counsel before you build a product on it.

• If your deployment can call an API, the licence question disappears and the comparison reverts to quality, price and features — where Flash TTS is ahead on two of the three.

• If you want open weights you can commercialise without a conversation, neither of these is that model, and the arena has other rows.

Where the Elo gap does and does not mean something

A 56-point gap is outside the ±16 and ±16 confidence intervals the board reports for these two rows, so the ranking is not noise in the way that, say, Flash TTS against Gemini 3.8 Flash-Lite TTS is noise. But an arena Elo measures one thing: which of two clips a listener preferred in a blind comparison. It is a preference signal, aggregated.

What it does not measure is everything a production pipeline cares about. It does not tell you how either model handles a 40-minute script without identity drift, how it pronounces your product's name, how it behaves when the input contains a phone number, or what its failure mode looks like when it is asked for something outside its training. Breeze TTS 2's 1,390 samples and Flash TTS's 1,999 are both large enough that the aggregate is stable, and both are small enough that no individual segment of your use case is represented in them.

The gap is also narrower than it looks once you account for what each model is being asked to do. Flash TTS is doing two speakers in one call, which is a harder task than the one Breeze TTS 2 is scored on. Whether that makes the 56 points an understatement or an overstatement depends on your use case, and neither vendor has published the ablation that would settle it.

What an hour of each costs

The two price figures are not directly comparable, and pretending otherwise is the most common error in write-ups of this matchup.

Google bills in tokens: $0.50 per million text tokens in and $9.00 per million audio tokens out through December 31, 2026, then $18.00. Batch and Flex run $0.25 and $4.50; Priority runs $0.90 and $16.00. Flash-Lite TTS, the sibling model, is $0.50 and $6.00, then $12.00.

BreezeBlue bills in characters: $34.00 per million characters through the hosted API.

Artificial Analysis normalises both onto a per-million-characters basis so they can share a board, and that is where $16.50 against $34.00 comes from. The normalisation is the board's, not either vendor's — Google does not publish a per-character rate for Flash TTS, and treating $16.50 as a quote rather than a conversion is how people end up surprised by an invoice.

Two consequences worth planning around:

• The Flash TTS rate doubles on January 1, 2027. Budgeted at the promotional number, the gap to Breeze TTS 2 narrows sharply; budgeted at $18.00, it nearly closes.

• Self-hosting Breeze TTS 2 does not make it free, it makes it a capital cost. A 3B model is small enough to run on modest hardware, but the arithmetic only wins at volume, and only for a use case the research licence permits.

Choosing, without pretending there is one answer

The decision tree is short.

Pick Gemini 3.8 Flash TTS if you need two-speaker dialogue, if you want a voice designed from a description and kept as an ID, if your pipeline can call an API, and if the December 31 rate change is something you can plan around. It is the better model on the independent board and the cheaper one on the bill, and the fact that it is closed is not a problem for a use case that was always going to be an API call.

Pick Breeze TTS 2 if data residency or on-premises inference is a hard requirement, if you need a 3B model you can actually download, or if you are doing research rather than shipping a product — which is precisely the case its licence is written for. Do not pick it for the open-weights label alone; read the licence first.

And if you are evaluating both, the cheap way to do that is to keep the choice reversible. OrcaRouter hosts neither of these models — the place to call Gemini 3.8 Flash TTS is Google's own API and the surfaces Google named, and the place to call Breeze TTS 2 is BreezeBlue's hosted endpoint. What a router is good for here is the rest of the stack: 200-plus models behind one key at provider list price with no markup, so a vendor price change like the January rate step is live the same day rather than at the next billing cycle, automatic failover so a path under evaluation does not have to be a path in production, and a routing DSL that composes models into a single call when one voice is not enough.

The one-line version: the leaderboard says Google, the licence says it depends, and the price says check your calendar before you commit to either.