Fish Audio Drama 3 voice API price set to jump fivefold after Nov 12

Fish Audio’s Drama 3 text-to-speech model, released on 13 October 2026, is about to leave its launch pricing: the API costs $10 per million characters until November 12 and $50 after that, according to an eesel.ai breakdown of Fish Audio’s launch post. That is a fivefold increase that, by eesel.ai’s reading, applies to every request after the deadline regardless of when a balance was topped up.

Drama 3 is Fish Audio’s directable voice model: it takes stage directions in square brackets, voices multi-character scenes in one pass, and ships as downloadable weights as well as through Fish Audio’s apps and API. For builders weighing it against hosted rivals, the price change, the self-hosting requirements and the licence terms all matter more this week than the benchmark headlines.

What Drama 3 is and what it adds

Hermes AI describes Drama 3 as a 35-billion-parameter text-to-speech model using a sparse mixture-of-experts (MoE) design, in which each token is routed through only a subset of the network; 3 billion parameters are active per token. The weights are published on Hugging Face.

The main selling point is control over delivery. Fish Audio describes Drama 3 as the first downloadable checkpoint that interprets bracketed stage directions like instructions to a voice actor, according to Hermes AI. Examples include tags such as [sighs] or [choked up]. OpenArt, which offers the model on its platform, gives a fuller picture of how that works in practice:

• A note such as [voice drops to a terrified whisper] can be placed before words, and the mood can change partway through a line.
• Individual words can be marked to be whispered, stressed, softened, sped up or slowed down, and sounds such as a sigh, laugh, breath or pause can be inserted at exact points.
• A whole scene with several characters can be voiced in one generation, with each character keeping its own voice and reacting to the previous line.
• The same voice carries across languages, and characters can switch language mid-scene.

Hermes AI reports support for up to 16 speakers in one pass, voice cloning, emotion changes within a sentence, and 83 languages for the downloadable checkpoint. OpenArt puts the language count at “more than 80”. Fish Audio’s own voice library pages now label community voices as “Powered by Fish Audio Drama 3”.

Under the hood

According to Hermes AI, one in every four attention layers uses full attention, while the others use linear attention via Gated DeltaNet, a mechanism that carries information across long sequences more cheaply. Audio is produced by a 10-codebook Codebook-MTP decoder: codec codebooks represent separate layers of audio as discrete tokens, and multi-token prediction lets the model emit several audio tokens per step. The MoE weights ship in 8-bit FP8 format and the rest in 16-bit bfloat16. Speakers are selected with vocabulary tokens from <|speaker:0|> through <|speaker:15|>, and long inputs are split automatically and can stream chunk by chunk.

Spec	Drama 3 (as reported)
Parameters	35B total, 3B active per token
Languages	83
Speakers per generation	Up to 16
Weight formats	FP8 (MoE weights), bfloat16 (rest)
Self-hosting hardware	NVIDIA GPU, 48 GB+ VRAM
Serving stacks	SGLang-Omni or vLLM-Omni
Released	13 October 2026

The November 12 price change, and conflicting figures

eesel.ai, which says it checked Fish Audio’s published pricing on 14 October 2026, reports two separate launch promotions that both run until November 12:

• API: $10 per million characters from October 13 to November 12, which eesel.ai calls 80% off the $50 list price. Because Fish Audio frames it as a date window, eesel.ai expects all requests after November 12 to bill at $50.
• Web and mobile plans: new Plus, Pro or Max plans get 50% off the first month on monthly billing, or the first year on annual billing. An annual Pro plan bought before the deadline costs $450 for twelve months instead of $900, per eesel.ai.

eesel.ai also reports that Fish Audio’s launch press release offers enterprises switching from another voice provider their first three months free.

Using Fish Audio’s own conversion of roughly 12 hours of English speech per million bytes, eesel.ai estimates the API cost at about $0.83 per hour of audio during the launch window and about $4.17 per hour afterwards. It cautions that vendors assume different reading speeds, so per-hour comparisons are only approximate and per-character comparisons are safer.

The figures do not line up across sources. Hermes AI reports a hosted price of $15 per million characters for Drama 3, which matches neither the $10 launch rate nor the $50 list price reported by eesel.ai. eesel.ai adds that Fish Audio’s official API pricing table did not list a drama-3 row at the time of its check, showing only s2.1-pro, s2.1-pro-free, s2-pro and the deprecated s1, so the launch post was the only place the Drama 3 price appeared. Builders budgeting beyond November 12 should confirm the current rate directly with Fish Audio.

What actually drives the bill

eesel.ai flags three details that change costs more than the headline rate:

1. Bracketed directions count as billed text, so heavily directed scripts cost more per spoken word.
2. Retakes are billed in full.
3. The older S2.1 Pro model is priced per UTF-8 byte, while Drama 3 is priced per character. For Chinese, Japanese and Korean, where one character takes several bytes, the Drama 3 premium nearly disappears.

The API is pay-as-you-go with “no subscription fees or monthly minimums,” per Fish Audio’s docs as quoted by eesel.ai.

How Drama 3 ranks on independent voice tests

Hermes AI reports results from Artificial Analysis, a third-party evaluator whose arena Elo scores come from pairwise listening tests. Its Provider Voice Arena uses each provider’s default voice; its Controlled Voice Arena gives every model the same cloned voice to reduce the effect of voice choice.

Benchmark (Artificial Analysis)	Score	Overall rank	Open-weight rank
Provider Voice Arena	1,278 Elo	No. 6	No. 1
Controlled Voice Arena	1,085 Elo	No. 12	No. 1
Pronunciation Robustness	94.6%	No. 1	No. 1

Figures as reported by Hermes AI.

On Provider Voice, Drama 3 finished 38 Elo points ahead of Drama 3 Preview and 50 ahead of BreezeBlue’s Breeze TTS 2; on Controlled Voice it narrowly led Ant Group’s Ming-3.0-tts at 1,068. The pronunciation test checks whether models expand abbreviations and units, choose context-dependent pronunciations, preserve codes and serial numbers, and say standalone terms correctly. Drama 3 placed first on expanding shorthand (96.6%) and second on contextual terms, behind Gemini 3.8 Flash TTS, but 13th on standalone terms at 91.8%, where Drama 3 Preview had scored 95.9%. Generation speed reached 66.6 characters per second, between Murf Speech Gen 2 at 63.0 and Eleven v4 Conversational at 66.8.

The picture is strong among open-weight models but mixed against the full field: five proprietary systems ranked above it on Provider Voice and eleven on Controlled Voice.

Self-hosting: hardware, speed and the licence

Local deployment requires an NVIDIA GPU with at least 48 GB of VRAM using the FP8 weights; CPU-only execution, Apple MPS and AMD ROCm are not supported in the current release, according to Hermes AI. Even with only 3 billion active parameters, self-hosters must hold the full checkpoint in GPU memory.

Fish Audio’s own serving tests on a single NVIDIA H200, as reported by Hermes AI, show a median time to first audio (TTFA, the delay before playback can start) of 88.9 ms for one stream, rising to 450.9 ms at 64 concurrent streams. The real-time factor (RTF, generation time divided by audio length, where below 1.0 means faster than playback) stays below 1.0 at 64 streams, at 0.801. Moving the vocoder to a second GPU raised throughput at 256 concurrent streams by roughly 72% and cut median TTFA by 37% in Fish Audio’s tests. These are the maker’s numbers, not independent measurements.

The weights are licensed under the Fish Audio Community License. Research use and commercial use below US$500,000 in annual revenue, including affiliates, carry no licence fee, and the licence bars reselling the model as a hosted text-to-speech API. Hermes AI notes that these commercial restrictions place Drama 3 in the open-weight category rather than fully open source.

What it means for builders

• Batch work has a deadline. eesel.ai’s practical advice is to render any backlog of scripts before November 12 and budget steady-state usage at $50 per million characters. After the deadline, eesel.ai calculates the API lands between the Pro and Plus web plans per hour, so the API is no longer automatically the cheaper route.
• Directed scripts cost more. Because bracket tags are billed text, products that generate many stage directions will see a higher effective rate than plain narration.
• CJK workloads compare differently. Per-character billing narrows the gap with the per-byte S2.1 Pro for Chinese, Japanese and Korean.
• Self-hosting is a real option for smaller companies. Teams under the US$500,000 revenue threshold with access to a 48 GB NVIDIA GPU can avoid per-character fees entirely, but cannot turn the model into a competing hosted API.
• Voice cloning carries consent questions. Drama 3 supports cloning; products built on it need their own consent and misuse safeguards.

What to watch next

The first thing to watch is whether Fish Audio’s official API pricing table adds a drama-3 entry, and at which price, given that eesel.ai found none on October 14 and the reported figures range from $10 to $15 to $50. Also worth tracking: whether later Artificial Analysis updates move Drama 3 against proprietary rivals, whether the weak standalone-terms score improves over the Preview’s result, and whether hardware support expands beyond NVIDIA GPUs with 48 GB or more.