Fish Audio Drama 3 vs Breeze TTS 2: The Open-Weights Comparison That Ends on a VRAM Figure

Fish Audio Drama 3 and Breeze TTS 2 are the two highest-ranked open-weights text-to-speech models on Artificial Analysis' Provider Voice Arena, 50 Elo apart, and they are not really competing for the same deployment. Drama 3 is a 35-billion-parameter mixture-of-experts that wants a single 48 GB card. Breeze TTS 2, from BreezeBlue, is a 3-billion-parameter dense model with published PyTorch inference code. That difference in footprint decides more purchasing questions than the Elo gap does, because the model you cannot run is not a model you can evaluate — and the two licences pull in opposite directions on top of it.

Same label, two different kinds of open weights

Both entries show a weights link on the board, and both would be described in a press release as open. Neither is open source, and the restrictions are not the same shape.

Breeze TTS 2's weights went up on Hugging Face on 2026-08-25 as BreezeBlue/Breeze-TTS-2, alongside the PyTorch inference code, under a research-and-non-commercial licence. That is a hard stop for commercial deployment without a separate arrangement, and it has not changed since. The company announced the model on 2026-08-16 and 2026-08-17 as a real-time voice model for interactive media, opened a hosted API at $34 per million characters, and published weights nine days later. BreezeBlue was founded by Yang Bin, a former MiniMax technical co-founder, on a $6 million seed round led by Yuanjing Capital and Redpoint China.

Fish Audio Drama 3 ships under the Fish Audio Community License v1.0, which is materially more permissive on the axis most teams care about: commercial use is free below US$500,000 in annual revenue for you and your affiliates, and the FAQ treats producing finished audio for clients as permitted. Above that threshold, or if you want to host Drama 3 as a text-to-speech, voice-cloning or inference service, you need a licence from Fish Audio. The licence also bans cloning a real person's voice without consent. S2 — the predecessor — remains under the stricter Research License, so the two Fish Audio generations cannot be treated as one legal category.

The practical read: Breeze TTS 2 is the easier model to experiment with and the harder one to ship commercially; Drama 3 is the reverse. If your evaluation is a research demo or an internal tool, Breeze's licence is not in your way and its 3B footprint makes the experiment cheap. If your evaluation ends in a product, the Fish Audio terms are the ones a lawyer will sign without a negotiation.

One detail on the Breeze model card is worth flagging because it is now out of date. The card's introduction still says the model "ranks #1 among open-weight models on the Artificial Analysis TTS leaderboard". That was true when it was written; the board today places Breeze TTS 2 thirteenth on provider voices and twenty-eighth on controlled voices, behind both Fish Audio entries. Model cards age faster than leaderboards, and a reader who checked only the card would come away with the wrong ranking.

The arena numbers, and why the controlled-voice gap is the bigger one

Read off Artificial Analysis on 2026-10-29:

• Provider Voice Arena — Fish Audio Drama 3 at rank 6 with an Elo of 1,278 and a board-recorded $15.00 per million characters; Breeze TTS 2 at rank 13 with 1,228 and $34.00. A 50-point gap, with the models ranked between them all closed systems.

• Controlled Voice Arena — Fish Audio Drama 3 at rank 12 with 1,085; Breeze TTS 2 at rank 28 with 1,021. A 64-point gap, and this is the board where every model is scored on the same fixed reference voices instead of on the voices its own vendor designed.

• The direction of the widening — Drama 3's advantage grows by 14 points when you take the vendor's own voice catalogue out of the comparison. That is the single most useful number in this matchup, because it says the Fish Audio lead is not an artefact of how nice its stock voices sound.

Fifty or sixty-four Elo is a real difference in blind preference tests and it is not a difference that survives being the only model in the comparison. Both are far below the closed frontier: Eleven v4 Turbo leads the provider board at 1,336 and HeyGen Voice leads the controlled board at 1,193. What the open-weights pair offers is a self-hostable floor, not a frontier.

What each one actually needs to run

This is where the comparison stops being abstract.

• Drama 3 — 35B total parameters, roughly 3B active per token, 8 of 256 routed experts plus one shared. Weights are about 40 GB across a shard set plus a separate codec file, and FP8 inference wants around 48 GB of VRAM. Fish Audio documents no CPU, Apple Silicon or ROCm path. On a single H200 the vendor reports 271 frames per second against 177 for S2, which is the payoff for the sparse architecture; that figure is vendor-reported and unreproduced outside Fish Audio.

• Breeze TTS 2 — 3B dense parameters. A BF16 checkpoint is roughly 6 GB, so it fits comfortably on a 24 GB card and plausibly on smaller ones with quantisation. Inference code is in the repository rather than implied. No published latency figure at the level of a single stream, but the model was designed and sold for real-time interactive media, which is a different requirement from high-fidelity narration.

• Tokenizer — Drama 3 keeps the Descript Audio Codec setup from S2: 44.1 kHz, ten codebooks, about 21.5 frames per second, fully causal. Breeze TTS 2's tokenizer is not something I can quote from a primary source, and a comparison that invented one would be false precision.

The consequence is blunt. A team with four 24 GB cards can run Breeze TTS 2 today and cannot run Drama 3 without sharding tricks Fish Audio does not document. A team with one H200-class card is choosing between a model that fits its hardware and a model that is 50 Elo better. Neither of those is the wrong answer; they are answers to different questions.

Cost at volume, in the unit Fish Audio actually bills

Fish Audio's documentation prices per million UTF-8 bytes, not per character, which changes the arithmetic for anyone not writing English. One million UTF-8 bytes is, per that same documentation, about 180,000 English words or twelve hours of speech. Drama 3 lists at $50.00 per million bytes — about $4.17 per finished audio hour — with a launch rate of $10.00, about $0.83 an hour, through 2026-11-12 09:00 PT. Breeze TTS 2 is $34.00 per million characters, and at the same rough conversion that is about $2.83 an hour.

So the ranking flips depending on the day you buy. During the promotion, Drama 3 is roughly a third of Breeze TTS 2's per-hour cost. After 2026-11-12, Drama 3 at list is roughly 47% more expensive per hour than Breeze — while being 50 to 64 Elo better on the arenas and materially better on pronunciation. And for CJK or Cyrillic text the byte-based billing pushes Drama 3 further up: a three-byte character costs three times what an English character does under that table, whereas a per-character vendor charges the same for both.

Two billing details from the same documentation are worth carrying into a budget model, because both inflate real spend quietly. Rate limits are on concurrent requests, not requests per second — 5 concurrent at the starter tier, 15 from $100 paid, 50 from $1,000 — and a request over the limit returns a 429 with no retry-after header, so a naive parallel renderer has to implement its own backoff. And requests that name a model id which is missing or not an exact match are served and billed as the default model, which on the TTS side means a misspelled Drama 3 header during the promo can bill you at S2.1 Pro's rate instead of the discount.

Putting both behind one pipeline

Neither Fish Audio Drama 3 nor Breeze TTS 2 is one of the models on OrcaRouter. Each is reached through its vendor's own API and, in the Fish Audio case, third-party platforms, and the honest version of this section says so instead of implying a route we do not have.

The orchestrating layer is a separate question, and it is where the two-model split earns its keep. A small studio that wants Drama 3 for narration and Breeze TTS 2 for real-time UI prompts is running two SDKs, two rate tables and two failure modes against one deadline. Routing the rest of the stack through one endpoint at provider list price with no markup, with automatic failover across upstreams, means the TTS vendor choice becomes a line in a config rather than a second contract — and the routing DSL can compose a model call with the audio step instead of bolting them together in application code.

If you are picking one today and the hardware is already bought, the decision is almost decided for you: 48 GB of VRAM is a different budget line from 6 GB, and Breeze's research-only licence is a different legal review from Drama 3's revenue-capped community licence. If you are picking one for quality within the open-weights class, the answer is Drama 3 on both arenas and on pronunciation, and the price argument only holds for the next two weeks.