Fish Audio Drama 3 Preview Tops the Open-Weights Speech Arena

Fish Audio Drama 3 Preview is the model Fish Audio published on October 13, 2026, and by the following morning it had taken the top spot among open-weights systems on the Artificial Analysis text-to-speech leaderboard. What makes that placement worth a second look is not the peak — it is the price attached to it. The entry carries a $15.00 per million characters list rate while every model ranked near it costs two to five times more, and the two models directly above it on the board, ElevenLabs Eleven v4 Turbo and ElevenLabs Eleven v4, are closed systems with per-character pricing that scales the wrong way for anyone producing long-form audio.

On the leaderboard's Provider Voice Arena, where each vendor's own voices compete against each other, Fish Audio Drama 3 Preview sits at rank 10 with an Elo of 1,243. That is behind the ElevenLabs pair at 1,330 and 1,328, behind Cartesia Sonic 3.6 at 1,286, behind Qwen-Audio-3.1-TTS-Plus at 1,300 and Qwen-Audio-3.0-TTS-Plus at 1,268, and behind Gemini 3.8 Flash TTS at 1,274. It is ahead of the second open-weights entry, Breeze TTS 2, which sits at 1,221 and costs $34.00 per million characters — a little over twice the Drama 3 Preview rate. Of the 97 models on that board, 17 are open weights, and Fish Audio's new preview is the highest-ranked of them.

Read the number with its error bars attached. The board's reported interval for the entry runs from 1,226.8 to 1,258.8, roughly 1,857 blind comparisons, across 8 arena voices. Against Breeze TTS 2 the gap is comfortable. Against Speechify Simba 3.2 at 1,245 and $6.60, it is not — that is a 2-point difference with overlapping intervals, and no one should be writing a purchase decision on it. The honest summary is that Drama 3 Preview put an open-weights model into the top ten of a board it previously had no business being in, and it did so at a mid-tier price.

What the model actually is

Fish Audio describes Drama 3 Preview as a 35-billion-parameter mixture-of-experts model with roughly 3 billion parameters active per token. It uses a hybrid attention stack — Gated DeltaNet combined with gated full attention — paired with a codebook-MTP decoder and a DAC-based tokenizer running at 44.1 kHz. Partial codebook fusion feeds only the first two codebooks back into the loop. Training ran through three pre-training stages, then instruction fine-tuning, then reinforcement learning against a single human-preference reward model with no rule-based reward shaping.

The throughput claim that comes with it is the part practitioners will care about most: Fish Audio reports 271 frames per second on a single H200, against 177 for the S2 model it succeeds, despite carrying far more parameters. This is a vendor figure, not a verified benchmark, and it will move with batching and sequence length — but the direction is consistent with a sparse architecture doing what sparse architectures are supposed to do.

The API surface is unchanged from the S2.1 Pro endpoint. Requests go to POST https://api.fish.audio/v1/tts, and a request that omits the model field or names one the service does not recognize falls back to s2.1-pro rather than erroring. That fallback is convenient and worth knowing about, because it means a typo in the model name produces plausible-sounding audio in the wrong voice rather than a 400.

Directorial control, dialogue in one pass, and single-word repair

Three capabilities separate this release from the previous generation, and all three are aimed at scripted production work rather than live conversation.

• Natural-language direction. Instead of a fixed tag vocabulary, the model accepts plain-description instructions such as [nervous, barely holding it together] or [warm and relaxed, like talking to an old friend], and it can change emotion or delivery mid-sentence without generating a new take. Paired inline tags like <whisper>…</whisper> and cues like [sigh] apply only to the text they wrap, so a line can shift register halfway through by construction rather than by luck.

• Multi-speaker dialogue in a single pass. Each character keeps its own voice inside one generation, which removes the splicing step that usually sits between a TTS model and a finished scene. For audio drama and game dialogue this is the difference between a pipeline and a research project.

• Audio inpainting. A segment can be re-selected, given a new instruction, and regenerated on its own, which is the workflow fix for the one mispronounced word that used to cost a full re-render. Fish Audio lists the interactive editing experience, the inpainting tool and an editing API as invite-only or coming soon, so the capability exists but the surface is not generally open yet.

Fish Audio also claims 83 languages with language consistency across long scripts and mid-scene switching, phoneme-level pronunciation control in English, Chinese, Japanese, Korean and several other languages, and S2-style voice cloning from reference clips of roughly 10 to 30 seconds. The launch demo used an Uncle Roger voice clone that the company says was made and used with permission.

The pronunciation claim, and where it comes from

Fish Audio's announcement states that Drama 3 Preview scores 93.2% overall on an Artificial Analysis Pronunciation Robustness leaderboard and ranks first there, ahead of ElevenLabs Eleven v4 and Eleven v4 Turbo. The sub-scores it publishes are 95.9% on names, brands and technical terms against 87.3% for Fish Audio S2.1 Pro; 95.3% on context-dependent words against 86.1%; 93.6% on numbers, dates and units against 92.7%; and 86.3% on codes, paths and email addresses against 79.4%.

Those numbers come from Fish Audio's own announcement. When I went looking for the board to read the per-model scores directly, the pronunciation-robustness paths returned 404s and the leaderboard surface I could load rendered the Elo ranking without a citable per-model pronunciation figure attached. Every figure in the paragraph above is therefore vendor-reported and should be labelled that way in any deck that uses it. The Elo numbers elsewhere in this article are different: those I read off the board, and they carry the board's own confidence intervals.

The underlying complaint the claim addresses is real regardless of whose number you trust. Pronunciation of proper nouns, product codes and file paths is the failure mode that pushes teams back to human voice talent, and it is the one most TTS vendors have been slowest to fix because it is not a prosody problem. A model that handles 86.3% of codes, paths and emails against a predecessor's 79.4% is still wrong roughly one time in seven.

What "open weights" means under this license

Fish Audio Drama 3 Preview ships under the Fish Audio Community License v1.0. Read the terms before you plan around them. The license is free for research and non-commercial use, and it is free for commercial use if you and your affiliates stay under US$500,000 in annual revenue. Above that threshold — or if you want to offer the model as a text-to-speech, voice-cloning or inference service, regardless of size — you need a license from Fish Audio. The S2 line remains under a separate Research License. Cloning a real person's voice without consent is prohibited outright.

The commercial threshold is corroborated independently: the Artificial Analysis board flags the entry as open weights without commercial-use rights, which is consistent with a revenue-gated license rather than an unrestricted one. If you are a small studio, open weights under this license is a genuine option. If you are a platform reselling inference, it is not open weights in the sense you were hoping for.

Pricing and the launch offer

The API list price is $50 per million characters, discounted 80% to $10 per million from October 13 through November 12, 2026. Creator credits match S2.1 Pro, and the new Plus, Pro and Max plans are half price through the same window. Artificial Analysis records the board price as $15.00 per million characters under default settings, which is the number to use when comparing against other entries on that board rather than against Fish Audio's own rate card.

The promotional window is the part with a real deadline on it. At $10 per million characters, Drama 3 Preview undercuts essentially every competitive model on the leaderboard by a wide margin. At $50, it is more expensive than Cartesia Sonic 3.6 at $49 and far above Qwen-Audio-3.0-TTS-Plus at $19.30. Anyone building a cost model on the discounted rate should build a second one on the list rate and check that the project still works.

Where this fits in a production stack

Scripted, render-once audio is the sweet spot: audiobooks, audio drama, game and animation dialogue, short-form video. The multi-speaker single-pass generation and the inpainting workflow both exist to serve that shape of work, and the pricing is per character rather than per minute, which favours long scripts. Interactive editing and the editing API are not generally available yet, so a workflow that depends on them is a workflow that depends on an invite.

The operational question after that is routing, and it is a harder one for speech than for text. Fish Audio Drama 3 Preview is not one of the models on OrcaRouter today. It is reached through the vendor's own API and several third-party platforms. The same is true of most of the models compared above — the speech leaderboard's top ten is a set of vendor-specific endpoints with vendor-specific authentication, and each one is its own integration and its own outage domain. OrcaRouter's ~200 models behind one key at provider list price with no markup covers text, vision and reasoning work, and its automatic failover exists precisely for the case where one provider's endpoint degrades mid-run. The practical pattern for a production speech path is to keep the speech vendor direct and keep everything it orchestrates — the summarisation, the translation pass, the pronunciation dictionary builder, the QA classifier — behind a single routed endpoint so a failover there does not take the render pipeline with it.

Bottom line

Fish Audio Drama 3 Preview is a real result rather than a launch-week artefact: rank 10 on the Provider Voice Arena at an Elo of 1,243, first among the board's 17 open-weights entries, at $15.00 per million characters against $34 for the next open-weights model down. The gaps to the top of the board are still large and the confidence intervals overlap with its nearest neighbours, so the open-weights crown is a starting position, not a lead. The 93.2% pronunciation figure is the vendor's own and could not be independently read off the board. The license has a US$500,000 revenue ceiling on commercial use that will rule it out for some teams entirely. And the $10 rate expires on November 12, 2026.

For a studio doing scripted audio at volume, the combination of multi-speaker single-pass generation, single-word inpainting and an open-weights license under the revenue threshold is the strongest argument here — stronger than the leaderboard position that will get the headlines.

Model choices like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.