Fish Audio Drama 3 vs Gemini 3.8 Flash TTS: Google Is the Cheap One Now, and That Reverses the Usual Argument

Price these two per minute of finished audio and the comparison stops resembling the one the market has been making all year. Google Gemini 3.8 Flash TTS bills at $9.00 per million audio tokens, and audio generates at 25 tokens per second — a little over two hundredths of a cent per second, which is about $0.81 per hour of speech. Fish Audio Drama 3 lists at $50.00 per million UTF-8 bytes, which the vendor's own documentation equates with roughly twelve hours of speech: about $4.17 per hour at list, and $0.83 per hour during the launch promotion that ends on 2026-11-12. So Google's flagship speech model is five times cheaper than the open-weights challenger once the discount expires, and roughly the same price while it lasts. The usual framing — a nimble open model undercutting a hyperscaler — is exactly backwards here, and understanding why is the whole decision.

The arithmetic, in enough detail to check

Google's unit is tokens, not characters, which is the reason the headline numbers looked unremarkable next to the text-per-character vendors. Gemini 3.8 Flash TTS outputs audio at 25 tokens per second of speech at $9.00 per million tokens through the end of 2026, and the model it supersedes — Gemini 3.1 Flash TTS Preview — was priced at $20.00 for the same million output tokens. Gemini 3.8 Flash-Lite TTS, shipped alongside on 2026-09-23, carries a shorter language list and a lower rate again, and Artificial Analysis records it at $11.00 per million characters for comparison purposes.

Fish Audio's unit is bytes, not characters, and its documentation states the conversion plainly: one million UTF-8 bytes is approximately 180,000 English words, or about twelve hours of speech. At $50.00 per million bytes that is $4.17 per hour before any overhead. At the promotional $10.00 it is $0.83. Direction brackets are billed as input text, so a heavily directed script costs 10-25% more than its word count suggests — and in this matchup every bracket is billed, because Drama 3's whole design is that you write the performance into the script.

• List price, per hour of speech — Gemini 3.8 Flash TTS about $0.81; Drama 3 about $4.17.

• Promotional price, per hour — Gemini 3.8 Flash TTS about $0.81; Drama 3 about $0.83. Effectively a tie.

• Unit — tokens per second of output on the Google side; UTF-8 bytes of input on the Fish Audio side. Both are honest units. They are not the same unit, and any "per million X" comparison that mixes them without converting is noise.

• The hidden line — a Drama 3 request whose model header is missing or misspelled does not fail, it falls back to s2.1-pro and bills at $15.00 per million bytes, which is a 50% surcharge over the promotional rate for a typo.

What Google's price actually buys

Being cheaper per hour is only interesting if the output is usable, and the independent numbers say it is, with a specific shape. Artificial Analysis places Google Gemini 3.8 Flash TTS fifth on the Provider Voice Arena at an Elo of 1,283 and a board-recorded $16.50 — above Fish Audio Drama 3's 1,278 — and twenty-fourth on the Controlled Voice Arena at 1,053, below Drama 3's 1,085. That split is the model in one line: it sounds excellent when Google's own voices are in play and it drops to the middle of the board when every model is judged on the same reference voices.

The pronunciation board tells a similarly narrow story. Gemini 3.8 Flash TTS leads all 31 scored models on the contextually appropriate category at 98.0%, and Drama 3 is at 95.3%. It is not the model's weak spot. Google's weak spot in this comparison is control: Gemini ships 30 prebuilt studio voices — Kore and Puck among them — with the full 130-language set and automatic language detection, and a voice library the documentation describes as "hundreds" in the Extended Voice Library while the marketing page says 2,000-plus. Those figures have never been reconciled from outside, and if a specific voice is load-bearing for your product, testing that voice is the only answer anyone can give you honestly.

Thirty usable voices with no cloning is fine for a support bot, a product narrator or a language-learning app. It is not fine for a drama pipeline, a game with twelve characters who need to be cast, or any client work where the voice is the deliverable. That is the gap Fish Audio fills, and it is a gap in kind rather than in degree.

What the open weights buy, and what they cost you to run

Fish Audio Drama 3 is a 35-billion-parameter mixture-of-experts with roughly 3 billion parameters active per token, initialised from Qwen3.5-35B-A3B, 8 of 256 routed experts plus one shared, forty layers in blocks of three Gated DeltaNet layers and one gated full-attention layer. Weights are published on Hugging Face as fishaudio/drama-3 — about 40 GB across a shard set plus a separate codec file — and FP8 inference wants around 48 GB of VRAM with no documented CPU or ROCm path. Serving is through SGLang-Omni, and the vendor reports 271 frames per second on a single H200 against 177 for S2.

That hardware line is the real trade. Google's price of $0.81 per hour includes no capital expenditure and no operations; your only fixed cost is the request. Drama 3's $0.83-per-hour promotional rate is an API rate, and running it yourself converts that into a 48 GB GPU, a serving stack and a person who owns it. For a catalogue of forty hours of audio a month, the API is obviously cheaper. For a few thousand hours, or for anything that has to run without a third party in the loop, the weights change the argument entirely — and they are the one thing Google is not selling at any price.

The licence matters here too, because it is what separates "open weights" from "free". Drama 3 is published under the Fish Audio Community License v1.0: free for research and non-commercial use, free commercially below US$500,000 in annual revenue for you and your affiliates, with a separate licence required above that threshold or to offer Drama 3 itself as a hosted speech or cloning service. Artificial Analysis marks it open weights without full commercial-use rights. Google's terms are the ordinary API terms, which is a different set of questions — data handling rather than redistribution.

Languages, and a number that is doing less than it looks

Google claims 130 languages with automatic detection from the text. Fish Audio claims 83, with a much stranger capability attached: directions can be written in any language and need not match the language of the script, and there is a documented language-switch tag that flips a single phrase into another language mid-line — marked experimental. Those are different products. Google's number is about how many markets you can enter with one integration. Fish Audio's is about how precisely one line can be produced, by a director who may not share the script's language.

For a team shipping a localised product, 130 beats 83 and the auto-detection is a real operational saving — no per-language routing, no per-language voice selection. For a team producing content in one or two languages for an audience that notices, 83 is not a constraint and the direction-in-any-language capability is the more valuable of the two features. Neither number should be read as a league table position; the count of languages a model claims has never correlated well with how good it is in the language you need.

Where each of them sits in a routed stack

This is one of the few articles in this series where the provider side is not fully honest to gloss over: OrcaRouter routes Google's earlier Gemini speech models — google/gemini-3.1-flash-tts-preview and google/gemini-2.5-flash-preview-tts are both in the catalogue today — but not Google Gemini 3.8 Flash TTS, and not Fish Audio Drama 3. The older Gemini TTS models are available through us at Google's list price with no markup; the two models this article is actually about are reached through their vendors' own APIs and, in Fish Audio's case, third-party platforms.

The useful consequence is that the larger part of a Google speech bill can already sit behind one key, and the failover layer is what makes the comparison cheap to run: put a canary of Drama 3 beside a Gemini 3.8 Flash TTS call, route the production path to whichever survives your acceptance test, and the per-segment choice lives in the routing DSL rather than in two SDKs. Model fusion is the other option this throws up — a panel that renders the same line through both and picks by a rule you define, which for narration means you can keep Drama 3's pronunciation on the lines with codes and Google's context reading on the prose.

The decision, stated plainly: for volume English with no cloning and no appetite for infrastructure, Gemini 3.8 Flash TTS at $0.81 per hour is the cheaper and better-sounding model on provider voices, and the argument for Drama 3 does not survive contact with the rate card. For anything that needs custom or cast voices, per-word direction, phoneme control, or the ability to run without a third party, Drama 3 is the one of the two that can do the job at all — at list price it costs five times Google's per hour, and during the promotion it costs the same, which makes the next two weeks the cheapest window anyone will get to find out whether the pronunciation advantage is worth building around.