Fish Audio ships two current models. OpenAudio S1 is the fast default; S2 Pro is the expressive flagship trained on 10M+ hours across 80+ languages. Both are MIT-licensed for self-hosting and available via the Fish Audio cloud.
S2 Pro ranked first in Fish Audio’s own 2026 blind-provider comparison. Artificial Analysis benchmarks agree: Fish Audio’s family currently leads on aggregate TTS quality ELO.
System Verdict
Pick Fish Audio if you need top-tier TTS quality without ElevenLabs pricing. S2 Pro is the strongest open-weight model in 2026, and self-hosting on a consumer GPU eliminates recurring cost entirely. S1 covers the fast, low-latency default; S2 Pro covers expressive narration and character work.
Skip it if the workflow is enterprise dubbing with lip-sync (use Resemble AI), if sub-100ms streaming latency is the hard constraint (use Cartesia), or if a no-code consumer UI matters more than raw quality (Speechify for reading, ElevenLabs for creator workflows).
Who pays which tier: Free for testing (7 minutes S2). Plus $11/mo for creators running ~200 minutes. Pro $75/mo for sustained 27-hour workloads. Max $749/mo for agency-scale 104-hour workloads. API at $15 per 1M characters for developers. Self-hosters pay only GPU cost.
Key Facts
| Flagship model | OpenAudio S2 Pro (dual-autoregressive, RL-aligned) |
| Fast model | OpenAudio S1 |
| Language coverage | 80+ languages including English, Chinese, Japanese, Korean, Spanish, French, German, Arabic, Hindi |
| License | MIT for open weights; weights on GitHub and Hugging Face |
| Self-hosting | Consumer GPUs with 8GB+ VRAM |
| Cloud free tier | 7 minutes S2 generation, 8K credits/mo |
| Cloud Plus | $11/mo, ~200 minutes S2, 250K credits |
| Cloud Pro | $75/mo, ~1,620 minutes (27 hours) S2, 2M credits, 3 team seats |
| Cloud Max | $749/mo, ~6,250 minutes (104 hours) S2, 25M credits, 10 team seats |
| API pricing | ~$15 per 1M characters · ~600-625 credits per minute |
| Blind-test rank (2026) | S2 Pro #1, S1 above every third-party provider |
Every data point above was verified against vendor sources on 2026-05-13. See Sources.
What it actually is
A single TTS stack served three ways. Weights on Hugging Face for self-hosters, a cloud platform at fish.audio for creators, and a REST API for developers.
S1 handles the default case: fast, low-latency, acceptable quality for most production voice work. S2 Pro handles expressive narration, multilingual output, and character voices where naturalness matters.
The moat is model quality plus license freedom. S2 Pro beats every commercial third-party provider in Fish Audio’s 2026 blind tests, and MIT weights mean no vendor lock-in. Artificial Analysis tracks Fish Audio’s family at the top of the current TTS ELO leaderboard.
When to pick Fish Audio
- You want the strongest open-weight TTS available in 2026. S2 Pro tops Fish Audio’s published blind-test rankings and Artificial Analysis’ ELO board.
- Self-hosting saves real money. High-volume inference runs at zero marginal cost after GPU setup.
- Multilingual coverage matters. 80+ languages on S2 Pro beats Voxtral’s 9 and matches ElevenLabs’ breadth.
- You need MIT-licensed weights. Forkable, fine-tunable, no training-data restrictions on commercial use.
- Expressive or character voices are the target. S2 Pro handles emotion and accent better than Cartesia or Voxtral in preference tests.
When to pick something else
- Enterprise dubbing with lip-sync: Resemble AI ships a full Localize pipeline and deepfake detection layer Fish Audio does not match.
- Sub-100ms streaming voice agents: Cartesia Sonic 3 hits 40-90ms time-to-first-audio; Fish Audio’s cloud latency lands higher.
- Cheapest commercial API: Voxtral at $0.016 per 1K chars (~$16/1M) is slightly cheaper and bundles STT.
- Polished consumer creator UI: ElevenLabs still wins on voice library breadth and creator workflow polish.
- Document-reading for personal use: Speechify solves the consumption case, not the production case.
Pricing
| Plan | Price | Included | Notes |
|---|---|---|---|
| Self-hosted | $0 | Unlimited | GPU with 8GB+ VRAM required |
| Free (cloud) | $0 | 7 min S2 / 8K credits | Testing only, non-commercial |
| Plus | $11/mo | ~200 min S2 / 250K credits | Best fit for most creators, 1 seat |
| Pro | $75/mo | ~27 hours S2 / 2M credits | Sustained production, 3 team seats |
| Max | $749/mo | ~104 hours S2 / 25M credits | Agency-scale, 10 team seats |
| API | $15 / 1M chars | Pay-as-you-go | ~600-625 credits per minute |
Prices verified 2026-05-13 via Fish Audio plan page and the Fish Audio 2026 blind-provider comparison.
Against the alternatives
| Fish Audio S2 Pro | ElevenLabs v3 | Voxtral | Cartesia Sonic 3 | |
|---|---|---|---|---|
| Blind-test quality | #1 in 2026 | Strong, second on aggregate | Wins vs ElevenLabs Flash v2.5 | Strong, tuned for real-time |
| Time-to-first-audio | Low, not sub-100ms | 200-400ms streaming | ~70ms multilingual | 40-90ms |
| Open weights | MIT | No | CC BY-NC 4.0 | No |
| Languages | 80+ | 30+ | 9 | 25+ |
| Commercial API | $15/1M chars | $30/1M chars | $16/1M chars | Credit-based |
| Best viewed as | Open-source quality leader | Creator platform default | Cheap multilingual API | Real-time agent specialist |
Failure modes
- Self-hosting requires ops. GPU management, model updates, and inference tuning fall on the deployer. Teams without GPU operations experience should stick to the cloud or API.
- Consumer UI trails ElevenLabs. The fish.audio dashboard covers the basics. It is not a full creator studio.
- Smaller voice library than ElevenLabs. Stock voices are limited; voice cloning fills the gap but needs clean input audio.
- Enterprise dubbing is not the product. No lip-sync, no automatic translation pipeline, no deepfake detection layer. Resemble ships that stack instead.
- Community tooling still catching up. Third-party plugins and wrappers exist but are thinner than ElevenLabs’ ecosystem.
- Credit math on the cloud plans. S1 and S2 consume credits at different rates. Heavy S2 users should price against the API or self-hosting before choosing Pro.
Recent changes
- 2026-05-13: Pricing reconfirmed live. Plus $11, Pro $75, Max $749. The Pro tier covers ~1,620 generation minutes (27 hours) per month with three team seats; the Max tier unlocks ~6,250 minutes (104 hours) and ten team seats for sustained agency or platform workloads. Credit math holds at roughly 600-625 credits per minute. Free tier remains 7 minutes plus 8K credits/mo, non-commercial.
- 2026-04-17: S2 Pro confirmed as flagship in Artificial Analysis’ aggregate TTS ELO leaderboard. API pricing unchanged at $15 per 1M characters.
Methodology
This page was produced by the aipedia.wiki editorial pipeline, an automated system that ingests vendor documentation, verifies pricing and model details against primary sources, and generates the editorial analysis you are reading. No individual human wrote this review. Scoring follows the four-dimension rubric at /about/scoring/ (Utility, Value, Moat, Longevity). Last verified 2026-05-13 against Fish Audio plan page, OpenAudio S2 page, Fish Audio 2026 provider comparison, and Artificial Analysis TTS leaderboard.
FAQ
Is Fish Audio free? Yes for self-hosting. OpenAudio S1 and S2 weights are MIT-licensed and run on consumer GPUs with 8GB+ VRAM. The cloud free tier gives 7 minutes of S2 generation for testing (Fish Audio blog).
What is the difference between S1 and S2? S1 is the fast default at low latency. S2 Pro is the expressive flagship trained on 10M+ hours across 80+ languages, and it ranked first in Fish Audio’s 2026 blind-provider comparison (S2 page).
How does Fish Audio compare to ElevenLabs? Fish Audio beats ElevenLabs on aggregate blind preference in 2026 Artificial Analysis benchmarks and the Fish Audio provider comparison. ElevenLabs still wins on creator-tool polish and voice library breadth.
Can I self-host for commercial use? Yes. MIT licensing permits commercial use of open weights with no royalty or training-data restrictions (GitHub).
Does Fish Audio support voice cloning? Yes. Both S1 and S2 support cloning from short reference samples. Quality improves with longer, cleaner reference audio.
Sources
- Fish Audio homepage: product positioning, current models
- Fish Audio plan page: Free, Plus, Pro, Max tiers; credits and minutes per tier
- OpenAudio S2 page: S2 Pro architecture and training data
- Fish Audio 2026 blind-provider comparison: S2 Pro #1 ranking
- Artificial Analysis: Fish Audio family: aggregate TTS ELO
- Fish Speech GitHub: MIT weights, self-hosting instructions
Related
- Category: AI Voice / TTS
- Comparisons: ElevenLabs vs Fish Audio, Cartesia vs Fish Audio, Fish Audio vs Voxtral, Fish Audio vs Resemble AI
