Fish Audio Drama 3 Now Holds Both Speech Arenas at Once — and Neither Number Is the One You Saw Last Week

The leaderboard carries two records for one product. Fish Audio Drama 3 sits sixth on the Artificial Analysis Provider Voice Arena at an Elo of 1,278 and twelfth on the Controlled Voice Arena at 1,085, and on both boards it is the highest-ranked open-weights model in the table. The record labelled Fish Audio Drama 3 Preview — the one that got the press when Fish Audio published it on 2026-10-13 — is still listed separately, at 1,240 and 1,063. That is a 38-point spread on one board and 22 on the other, for what the vendor's own product page calls a single model under two names. If you have quoted a Drama 3 Elo in the last two weeks, this is the piece that tells you which of the two you were quoting.

What the boards actually say today

Both figures were read off Artificial Analysis on 2026-10-29, and both come with the same board-recorded $15.00 per million characters, so the 38-point gap is not a price change — it is two evaluations of two archived states of the same model. The provider-voice standings above Drama 3 are all closed systems: Eleven v4 Turbo at 1,336 and $40.00, Eleven v4 at 1,334 and $80.00, Qwen-Audio-3.1-TTS-Plus at 1,305 and $19.30, Cartesia Sonic 3.6 at 1,296 and $49.00, and Google Gemini 3.8 Flash TTS at 1,283 and $16.50. Every one of them costs more per million characters than the Fish Audio entry ranked immediately below them.

The controlled-voice board is where the two Drama 3 records separate most sharply in rank, and where the open-weights claim is easiest to check, because all models there are scored on the same fixed reference voices rather than on each vendor's own catalogue. Full spread for the open-weights entries as of today:

• Provider Voice — Fish Audio Drama 3 1,278 (#6); Fish Audio Drama 3 Preview 1,240 (#12); Breeze TTS 2 1,228 (#13); Voxtral 3 TTS 1,223 (#15).

• Controlled Voice — Fish Audio Drama 3 1,085 (#12); Ming-3.0-tts 1,068 (#17); Fish Audio Drama 3 Preview 1,063 (#20); Breeze TTS 2 1,021 (#28).

• The gap that matters — on controlled voices, Drama 3 is still behind Eleven v4 (1,156, #4), Inworld Realtime TTS-2 (1,146, #7) and Google Gemini 3.8 Flash TTS (1,053, #24) is below it, but the closed systems at the top of that board are not four points away, they are seventy. HeyGen Voice leads it at 1,193.

Two things follow. First, "first among open weights" is true on both arenas today and it is not a tie on either. Second, the margin on the controlled board — 17 Elo over Ming-3.0-tts at a fifth of the parameter count — is inside the range where a few thousand votes can move the order, so treat it as a lead rather than as a lock.

The pronunciation score is a shared first, not an outright one

Fish Audio's announcement quotes 93.2% on Artificial Analysis' English Pronunciation Robustness benchmark, against 88.1% for its previous S2.1 Pro, and describes it as the top score on the leaderboard. That was accurate when the score landed on 2026-10-20, with a qualification the marketing copy leaves out: Ant Group's Ming-3.0-tts reached the same 93.2%, so the board shows a tie for first among 31 scored models rather than a sole leader. Eleven v4 sits at 91.7% and Eleven v4 Turbo at 90.1%.

The aggregate hides a real trade. Drama 3's strength is proper nouns and context-dependent words — 95.9% on standalone terms such as brands, places and technical names, and 95.3% on words whose reading depends on the sentence. Its weakness is anything that must be reproduced character for character: 86.3% on codes, paths and email addresses, which is where Ming reaches 90.6%. The one category Drama 3 loses outright is expanding shorthand — numbers, dates and units — at 93.6% against Eleven v4's 94.1%. If your pipeline reads order numbers, part codes or tracking strings aloud, the headline 93.2% is the wrong number to plan against; the 86.3% is the right one.

What shipped on 2026-10-13, and what sixteen days changes

None of this is a launch, and it should not be reported as one. Fish Audio put the model into its web app, its iOS and Android apps and its API on 2026-10-13, with the API model id "drama-3". What is new in the last fortnight is the independent scoring of the retail model, not the model's existence. The practical consequences of that distinction are worth spelling out, because the earlier coverage of this model — including on this blog — was written against the Preview data.

The architecture has not changed between the two records: a 35-billion-parameter mixture-of-experts initialised from Qwen3.5-35B-A3B, with roughly 3 billion parameters active per token and 8 of 256 routed experts plus one shared expert firing. Forty layers arranged in repeating blocks of three Gated DeltaNet layers and one gated full-attention layer, so only one layer in four keeps a KV cache. The decoder is a codebook-MTP design — four shared layers feeding nine full-width layers, one per acoustic codebook — and only codebooks 0 and 1 are fed back into the backbone. The tokenizer is the same Descript Audio Codec setup S2 used: 44.1 kHz, ten codebooks, about 21.5 frames per second, fully causal. One small inconsistency is worth noting because it will confuse anyone comparing the two sources: Fish Audio's announcement says 35 billion parameters, and the Hugging Face model card reports a model size of 36B. The difference is likely a rounding or embedding-count convention rather than a different checkpoint, but the two numbers are what you will find in each place.

Throughput on a single H200 is a vendor figure of 271 frames per second against 177 for S2, which is what a 35B MoE with a 3B active path is supposed to buy you. Serving is through SGLang-Omni. The weights are on Hugging Face as fishaudio/drama-3 — around 40 GB across a set of safetensors shards plus a separate codec file — and FP8 inference wants roughly 48 GB of VRAM, with no CPU or ROCm path documented. That last constraint is the one that decides most self-hosting questions, and it is the subject of the Breeze TTS 2 comparison rather than this one.

Three prices, and the card is not the one you should plan against

Here the sourcing has to be explicit, because three numbers are in circulation and two of them are not what they appear. Artificial Analysis records $15.00 per million characters for both Drama 3 entries. Fish Audio's own announcement page states a list price of $50.00 per million characters with a launch rate of $10.00, 80% off, through 2026-11-12. And Fish Audio's developer documentation bills in a third unit entirely: per million UTF-8 bytes, with drama-3 at $50.00, s2.1-pro at $15.00, and the launch rate at $10.00 until 2026-11-12 09:00 PT. English runs close to one byte per character, so the byte and character figures are roughly interchangeable for English text and diverge sharply for scripts that cost more than a byte a glyph.

The $15.00 on the leaderboard is worth a second look rather than a shrug, because it is exactly the list rate of s2.1-pro, Fish Audio's previous model, and not the rate of the one being scored. Either the board recorded a blended or carried-over figure, or it priced the model at a rate Fish Audio does not publish for it. The number to budget on is the one in Fish Audio's own pricing table: $50.00 per million UTF-8 bytes at list, which the provider's documentation equates with about 180,000 English words, or roughly twelve hours of speech — call it $4.17 per finished audio hour. A launch discount that expires in two weeks is not a rate card; at $10.00 the same twelve hours costs about $0.83.

Licence: open weights, not open source, and one clause that decides a lot

The weights are published under the Fish Audio Community License v1.0. Research and non-commercial use is free. Commercial use is free if you and your affiliates stay under US$500,000 in annual revenue; above that, or if you want to offer Drama 3 as a text-to-speech, voice-cloning or inference service, you need a separate licence from Fish Audio. The licence also prohibits cloning a real person's voice without consent, and Fish Audio's FAQ treats "making finished audio for clients" as permitted, which is the clause that matters for agencies and post houses. S2 remains under the stricter Research License, so the two generations are not interchangeable in a commercial deployment.

Artificial Analysis lists the model as open weights without full commercial-use rights. That is the honest description: you can download and run it, you cannot necessarily sell access to it. Anyone planning a hosted product on top of these weights should read the service clause before they build the billing page, not after.

Where this leaves a production pipeline

The reason the two-arena result is worth a post rather than a footnote is that it changes what a team can assume. Sixteen days ago the defensible statement was that an open-weights model had entered the top ten of a preference board and claimed a pronunciation lead on its own evaluation. Today the independent verdict is narrower and stronger: it is the best open-weights model on both arenas, it is a shared first on pronunciation rather than an outright one, and it is doing that at a mid-tier rate with weights anyone can download.

Drama 3 is not one of the models on OrcaRouter. It is reached through Fish Audio's own API and several third-party platforms, and a piece that pretended otherwise would be lying to you about the one thing this blog is for. What we do is the other half of the problem: the ~200 models behind a single key at provider list price with no markup, automatic failover, the routing DSL and model fusion are how a pipeline that already calls other vendors absorbs a model like this one without a second contract and a second SDK. If Drama 3 becomes the narration voice on a render queue and the queue has a deadline, the value of routing everything else through one endpoint is that a retry path already exists.

What to watch over the next fortnight, in order of how much it would change the picture: whether the 22-point gap between the two Drama 3 records closes or widens as more votes land on the retail entry; whether Ming-3.0-tts pulls ahead in pronunciation once its 21st-place proper-noun score is re-measured with a larger span count; and whether anyone outside Fish Audio reproduces the 271 frames per second figure on hardware that is not an H200. The arena numbers are settled enough to quote. The throughput number is still a vendor claim, and the licence is still the thing most likely to decide whether a team can use this at all.