New releases

Just arrived.

  1. NewNemotron 3.5 Super VL 120B A12BText and chat · NVIDIA
  2. NewMuse Spark OpenText and chat · Meta
  3. NewMuse WonderText and chat · Meta
  4. NewQwen-Image 2.1 TurboImages · Alibaba (Qwen)
  5. NewQwen-Image 2.1 Turbo (edit)Edit and retouch · Alibaba (Qwen)
  6. NewClaude Haiku 5.5Text and chat · Anthropic
  7. NewNano Banana 2.1Images · Google DeepMind
  8. NewNano Banana 2.1 (edit)Edit and retouch · Google DeepMind
  9. NewQwen-Image 2.1 ProImages · Alibaba (Qwen)
  10. NewQwen-Image 2.1 Pro (edit)Edit and retouch · Alibaba (Qwen)

Prices

What each kind of work costs.

Every price, model by model
Prices in US dollars, by kind of work
Kind of workCheapestSub Rosa’s defaultBest measured
Text and chatQwen 2.5 7B E2EE$0.0584 per million tokens writtenGLM 5.2$1.978 per million tokens writtenClaude Opus 5.5$8.992 per million tokens written
TranscriptionxAI Speech to Text v1$0.048 per hour of audioParakeet ASR$0.162 per hour of audioElevenLabs Scribe V2$0.27 per hour of audio
ImagesChroma$0.0045 per imageVenice SD35$0.0045 per imageGPT Image 2.5 Sunburst$0.0315 per image
Edit and retouchQwen Image 2.1 Turbo$0.009 per editIdeogram V4.5$0.117 per editGPT Image 2.5 Sunburst$0.036 per edit
VideoLongcat Distilled$0.0081 per second of videoKling 2.5 Turbo Pro$0.0351 per second of videoMiniMax H3 Max$0.0279 per second of video
VoiceKokoro Text to Speech$0.00157 per 1,000 charactersKokoro Text to Speech$0.00157 per 1,000 charactersElevenLabs TTS v4 Turbo$0.0207 per 1,000 characters
MusicMiniMax Music 2.0$0.018 per trackACE-Step 1.5$0.0135 per minute of musicNot measured
Sound effectsMMAudio V2$0.0041 per 10 seconds of soundNo defaultNot measured

The whole catalog

Every family.

Kind of work

100 families

Text and chat

Text and chat · Z.ai (Zhipu AI)GLM

Sub Rosa's default chat model: an open model from Z.ai with strong reasoning and tool use, in private mode.

  • Latest : GLM 5.3 Flash, August 26, 2026
  • #13 of 90 on Artificial Analysis Intelligence Index (45 points)
Private, some versions anonymized Moderate cost13 models
Text and chat · Moonshot AIKimi

Sub Rosa's suggested alternate: an open model that excels at long tool-driven tasks and reads images, in private mode.

  • Latest : Kimi K3, July 16, 2026
  • #15 of 90 on Artificial Analysis Intelligence Index (44 points)
Private Moderate cost7 models
Text and chat · DeepSeekDeepSeek

Low-cost open models with 1M context and good reasoning; V4.1 Flash adds image reading. Pick it to do a lot for very little.

  • Latest : DeepSeek V4.1 Flash, September 10, 2026
  • #24 of 90 on Artificial Analysis Intelligence Index (39 points)
Private Low cost8 models
Text and chat · AnthropicClaude

Anthropic's models, strong at writing, code and long agent work, with vision and 1M context, from the small Haiku to the top Fable. Anonymized, and mostly on the dear side.

  • Latest : Claude Haiku 5.5, October 7, 2026
  • #1 of 90 on Artificial Analysis Intelligence Index (58 points)
Anonymized High cost16 models
Text and chat · OpenAIGPT

OpenAI's models: Astra for the hardest work, Sol for coding and agents, Luna for cheap volume. Anonymized.

  • Latest : GPT-6.1 Sol, September 2026
  • #4 of 90 on Artificial Analysis Intelligence Index (53 points)
Anonymized High cost21 models
Text and chat · OpenAIGPT OSS

OpenAI's open-weight model, run in private mode: cheap, quick, with adjustable reasoning. Good for everyday questions.

  • Latest : gpt-oss-120b, August 2025
  • #79 of 90 on Artificial Analysis Intelligence Index (12 points)
Private Low cost2 models
Text and chat · GoogleGemini

Google's Gemini models: Argon, the new flagship, and the fast Flash line. 1M context, and they take images, audio and video as input. Anonymized.

  • Latest : Gemini 4 Argon, September 30, 2026
  • #5 of 90 on Artificial Analysis Intelligence Index (53 points)
Anonymized High cost8 models
Text and chat · xAIGrok

xAI's Grok models, in private mode, with vision, very long context and search on X. Grok 4.7 is the newest.

  • Latest : Grok 4.7, September 21, 2026
  • #12 of 90 on Artificial Analysis Intelligence Index (46 points)
Private Moderate cost7 models
Text and chat · Alibaba (Qwen)Qwen Max, Plus and Flash

Alibaba's hosted Qwen models, with 1M context and image and video input. Flash is cheap and fast. Anonymized.

  • Latest : Qwen3.8 Max, September 2026
  • #14 of 90 on Artificial Analysis Intelligence Index (45 points)
Anonymized Low cost6 models
Text and chat · Alibaba (Qwen)Qwen open models

Qwen models with open weights, run in private mode. Qwen 3.8 27B reads images and video at a low price.

  • Latest : Qwen3.8 27B, August 14, 2026
  • #23 of 90 on Artificial Analysis Intelligence Index (40 points)
Private Moderate cost15 models
Text and chat · Google DeepMindGemma

Google's small open models, run in private mode: very cheap, they read images and video and call tools.

  • Latest : Gemma 4, April 2026
  • #71 of 90 on Artificial Analysis Intelligence Index (17 points)
Private Low cost3 models
Text and chat · Meta (Hermes 3: Nous Research)Llama and Hermes 3

Meta's older open Llama models and the Hermes 3 fine-tune, in private mode. Newer families do better for the price.

  • Latest : Llama 3.3 70B, December 6, 2024
  • #86 of 90 on Artificial Analysis Intelligence Index (8 points)
Private Moderate cost3 models
Text and chat · Mistral AIMistral Small

A compact 24B model from Mistral AI, in private mode: very cheap, reads images and calls tools.

  • Latest : Mistral Small 3.2, June 20, 2025
  • #87 of 90 on Artificial Analysis Intelligence Index (8 points)
Private Low cost1 model
Text and chat · MiniMaxMiniMax

MiniMax's open models, tuned for coding and agent work, in private mode at a low price.

  • Latest : MiniMax M3, June 1, 2026
  • #61 of 90 on Artificial Analysis Intelligence Index (23 points)
Private Low cost3 models
Text and chat · InceptionMercury

Diffusion language models that write many tokens at once: very fast replies, with reasoning and tools, at a low price.

  • Latest : Mercury 2.5, October 1, 2026
  • #76 of 90 on Artificial Analysis Intelligence Index (14 points)
Anonymized Low cost2 models
Text and chat · XiaomiXiaomi MiMo

Xiaomi's open models that understand text, images, audio and video, with 1M context, in private mode and very cheap.

  • Latest : MiMo-V2.6-Flash, September 21, 2026
  • #32 of 90 on Artificial Analysis Intelligence Index (38 points)
Private Low cost2 models
Text and chat · NVIDIANemotron

NVIDIA's open models for agent work: Super VL reads images and video, Ultra handles long tool-driven tasks, Nano is small and cheap. Private mode.

  • Latest : Nemotron 3.5 Super VL 120B A12B, October 15, 2026
  • #52 of 90 on Artificial Analysis Intelligence Index (26 points)
Private Low cost3 models
Text and chat · ByteDanceSeed

ByteDance's multimodal model for coding and long agent tasks, reading images and video. Anonymized.

  • Latest : Seed 2.1 Turbo, August 2026
Anonymized Moderate cost1 model
Text and chat · Thinking Machines LabInkling

The first open model from Thinking Machines Lab: general purpose, reads images and audio, 512K context, private mode.

  • Latest : Inkling, July 15, 2026
  • #56 of 90 on Artificial Analysis Intelligence Index (25 points)
Private Moderate cost1 model
Text and chat · Venice.ai and Cognitive Computations (Dolphin)Venice Uncensored

Uncensored models for open creative writing and role play, based on Mistral Small 24B, in private mode.

  • Latest : Venice Uncensored 1.2, April 2026
Private Low cost2 models
Text and chat · AionLabsAion

Storytelling systems where several models write each reply together, for interactive fiction with mature themes.

  • Latest : Aion 3.5, September 2026
Anonymized Moderate cost4 models
Text and chat · Community fine-tunes (Abliteration AI, Olafangensan, others)Uncensored community models

Open models altered by the community to remove refusals. Pick one for unfiltered creative work.

  • Latest : Abliterated Large V2, October 2026
Private, some versions anonymized Low cost4 models
Text and chat · TencentHy

Tencent’s Hy4 preview: an open 770B mixture-of-experts model with 1M context, tuned for code, office work and agents.

  • Latest : Hy4 preview, August 2026
Private Moderate cost1 model
Text and chat · MetaMuse

Meta’s text models: Muse Wonder is its first frontier model, reading text, images, audio and video over 1M tokens, and Muse Spark Open is the open-weight one.

  • Latest : Muse Spark Open, October 15, 2026
  • #7 of 90 on Artificial Analysis Intelligence Index (52 points)
Anonymized, some versions private High cost2 models

Images

Images · Google DeepMindNano Banana

Google's Gemini image models: strong prompt understanding and real-world knowledge. Pick Nano Banana 2.1 as a reliable all-rounder.

  • Latest : Nano Banana 2.1, October 6, 2026
  • #4 of 37 on Artificial Analysis Text to Image Arena (1,172 Elo)
Anonymized Moderate cost4 models
Images · OpenAIGPT Image

OpenAI's image models, known for readable text and precise instruction following. Start with GPT Image 2.5 Flare.

  • Latest : GPT Image 2.5 Flare and Sunburst, September 8, 2026
  • #1 of 37 on Artificial Analysis Text to Image Arena (1,197 Elo)
Anonymized Moderate cost4 models
Images · Black Forest LabsFLUX

Black Forest Labs' photorealistic models. FLUX.2 Pro is what the Sub Rosa agent uses for detailed, realistic images.

  • Latest : FLUX 3 Image, July 23, 2026
  • #20 of 37 on Artificial Analysis Text to Image Arena (1,020 Elo)
Anonymized Low cost3 models
Images · Alibaba (Qwen)Qwen Image

Alibaba's models for images full of text: posters, menus, infographics. Qwen Image 3 stays the pick; 2.1 Turbo is the cheapest.

  • Latest : Qwen-Image 2.1 Turbo, October 9, 2026
  • #11 of 37 on Artificial Analysis Text to Image Arena (1,090 Elo)
Anonymized Low cost7 models
Images · ByteDance SeedSeedream

ByteDance's models for consistent subjects and clean typography across a series. Seedream 5 Pro is the flagship.

  • Latest : Seedream 5.0 Pro, July 8, 2026
  • #12 of 37 on Artificial Analysis Text to Image Arena (1,081 Elo)
Anonymized Low cost3 models
Images · IdeogramIdeogram

A model built around typography: posters, logos, covers where the words must be exactly right. Ideogram 4.5 is the current one.

  • Latest : Ideogram 4.5, September 2026
  • #23 of 37 on Artificial Analysis Text to Image Arena (1,013 Elo)
Anonymized High cost2 models
Images · RecraftRecraft

A design-minded model: balanced composition, cohesive color, structured text. Pick it for marketing visuals and brand work.

  • Latest : Recraft V4 and V4 Pro, February 17, 2026
  • #30 of 37 on Artificial Analysis Text to Image Arena (987 Elo)
Anonymized Low cost2 models
Images · xAIGrok Imagine

xAI's image models, run in private mode here. Grok Imagine 2.0 is the newest, a good choice when privacy matters.

  • Latest : Grok Imagine Image 2.0, August 2026
  • #5 of 37 on Artificial Analysis Text to Image Arena (1,155 Elo)
Private Moderate cost3 models
Images · TencentHunyuan Image

Tencent's large open-weight model, run in private mode. Pick it for long, knowledge-heavy prompts in English or Chinese.

  • Latest : HunyuanImage 3.0, September 28, 2025
  • #34 of 37 on Artificial Analysis Text to Image Arena (944 Elo)
Private Moderate cost1 model
Images · KreaKrea 2

Krea's own model, built for aesthetics and expressive styles rather than literal prompt following. Turbo is fast and runs in private mode.

  • Latest : Krea 2 Turbo, June 3, 2026
  • #19 of 37 on Artificial Analysis Text to Image Arena (1,026 Elo)
Private, some versions anonymized Low cost3 models
Images · Luma AILuma Uni-1

Luma's model that reasons before drawing, good at tricky spatial instructions and styles. Pick it when other models miss the scene.

  • Latest : Luma Uni-1 and Uni-1 Max, May 2026
  • #22 of 37 on Artificial Analysis Text to Image Arena (1,014 Elo)
Anonymized Low cost2 models
Images · Alibaba (Wan)Wan 2.7 Image

Alibaba Wan's image models with a thinking mode for complex scenes. Use the standard one for drafts, Pro for final product shots.

  • Latest : Wan 2.7 and 2.7 Pro, April 2026
  • #29 of 37 on Artificial Analysis Text to Image Arena (990 Elo)
Anonymized Low cost2 models
Images · ImagineArtImagineArt

ImagineArt's flagship, tuned for realism, mood and composition. Pick it for posters, product visuals and cinematic stills.

Anonymized Low cost1 model
Images · MetaMuse Image

Meta's first image model (July 2026), with clean in-image text. At about 1 credit, a cheap way to get readable graphics.

  • Latest : Muse Image, July 7, 2026
  • #7 of 37 on Artificial Analysis Text to Image Arena (1,115 Elo)
Anonymized Low cost1 model
Images · lodestones (open source)Chroma

An open model built from FLUX.1, run in private mode for about 0.5 credit. The Sub Rosa agent uses it as a versatile option.

  • Latest : Chroma1-HD, August 2025
Private Low cost1 model
Images · Stability AIStable Diffusion 3.5

Stability AI's SD 3.5 Large, private and about 0.5 credit. It is the Sub Rosa agent's default when you don't name a model.

  • Latest : Stable Diffusion 3.5 Large, October 22, 2024
  • #37 of 37 on Artificial Analysis Text to Image Arena (840 Elo)
Private Low cost1 model
Images · Alibaba (Tongyi Lab)Z-Image Turbo

A small, fast open model with good photorealism, private and about 0.5 credit. Pick it for quick drafts that still look real.

  • Latest : Z-Image-Turbo, November 2025
  • #35 of 37 on Artificial Analysis Text to Image Arena (941 Elo)
Private Low cost1 model
Images · Community fine-tune of Illustrious XLAnime (WAI Illustrious)

A popular anime fine-tune of Illustrious XL, private and about 0.5 credit. Pick it for anime characters and stylized art.

Private Low cost1 model
Images · Community fine-tune (SDXL)Lustify

SDXL fine-tunes for adult content, uncensored, private and about 0.5 credit. Use v8, the most recent.

Private Low cost3 models
Images · VeniceUpscaler

A utility model rather than a generator: the upscaler enlarges an image 2x to 4x, with an optional enhance pass, in private mode.

Private1 model

Edit and retouch

Edit and retouch · Alibaba (Qwen)Qwen Image (edit)

Instruction-following edits that keep the rest of the image intact. Qwen Image 2 is the Studio's first "Automatic" choice.

  • Latest : Qwen-Image 2.1 Turbo (edit), October 9, 2026
  • #12 of 26 on Artificial Analysis Image Editing Arena (1,075 Elo)
Anonymized Low cost6 models
Edit and retouch · ByteDance SeedSeedream (edit)

ByteDance's editors keep lighting, color and the subject's look while changing what you ask. Seedream 5 Lite is the second "Automatic" choice.

  • Latest : Seedream 5.0 Pro (edit), July 8, 2026
  • #8 of 26 on Artificial Analysis Image Editing Arena (1,107 Elo)
Anonymized Low cost3 models
Edit and retouch · Google DeepMindNano Banana (edit)

Google's conversational editors: describe the change in plain words. Nano Banana 2 is the fourth "Automatic" choice.

  • Latest : Nano Banana 2.1 (edit), October 6, 2026
  • #3 of 26 on Artificial Analysis Image Editing Arena (1,152 Elo)
Anonymized Moderate cost4 models
Edit and retouch · IdeogramIdeogram 4.5 (retouch)

Ideogram's precise editor, the one the Retouch tab opens with: paint a zone, describe the change, the rest of the pixels stay put.

  • Latest : Ideogram 4.5 (retouch), September 2026
  • #16 of 26 on Artificial Analysis Image Editing Arena (1,064 Elo)
Anonymized High cost1 model
Edit and retouch · OpenAIGPT Image (edit)

OpenAI's editors for localized changes and text fixes. Sunburst holds intricate detail best; Flare is quicker at the same price.

  • Latest : GPT Image 2.5 Flare and Sunburst (edit), September 8, 2026
  • #1 of 26 on Artificial Analysis Image Editing Arena (1,182 Elo)
Anonymized Moderate cost4 models
Edit and retouch · Black Forest LabsFLUX (edit)

Photoreal edits that keep faces, outfits and products recognizable. FLUX 3 is the newer of the two, at the same price.

  • Latest : FLUX 3 Image (edit), July 23, 2026
  • #25 of 26 on Artificial Analysis Image Editing Arena (995 Elo)
Anonymized Moderate cost2 models
Edit and retouch · xAIGrok Imagine (edit)

xAI's editors in private mode: the pick when the photo you edit should not be kept by anyone. Version 2.0 is the most recent.

  • Latest : Grok Imagine Image 2.0 (edit), August 2026
  • #7 of 26 on Artificial Analysis Image Editing Arena (1,107 Elo)
Private Moderate cost3 models
Edit and retouch · Luma AILuma Uni-1 (edit)

Luma's reasoning editor, good at camera angle, lighting and style changes on a single image.

  • Latest : Luma Uni-1 and Uni-1 Max (edit), May 2026
  • #14 of 26 on Artificial Analysis Image Editing Arena (1,070 Elo)
Anonymized Low cost2 models
Edit and retouch · MetaMuse Image (edit)

Meta's editor at about 1 credit: the cheapest way to try a change or blend a few photos.

  • Latest : Muse Image (edit), July 7, 2026
  • #5 of 26 on Artificial Analysis Image Editing Arena (1,118 Elo)
Anonymized Low cost1 model
Edit and retouch · Alibaba (Wan)Wan 2.7 Pro (edit)

Alibaba Wan's professional editor for complex, multi-element changes and restyling.

  • Latest : Wan 2.7 Pro (edit), April 2026
  • #21 of 26 on Artificial Analysis Image Editing Arena (1,042 Elo)
Anonymized Moderate cost1 model
Edit and retouch · Xiaohongshu (FireRedTeam)FireRed Image Edit

An open editing model run in private mode, strong on portraits, old photo repair and stylized text. About 2 credits.

  • Latest : FireRed-Image-Edit 1.0, February 2026
Private Low cost1 model
Edit and retouch · Qwen-based (uncensored variant)Qwen Edit Uncensored

A Qwen-based editor without content restrictions, run in private mode for about 2 credits.

Private Low cost1 model

Video

Video · KuaishouKling V3

Kuaishou's 2026 generation: text or image to video with sound and multi-shot prompts. Pick it as your all-round cinematic choice.

  • Latest : Kling 3.0, February 5, 2026
  • #12 of 22 on Artificial Analysis Image to Video Arena (1,055 Elo)
Anonymized Moderate cost12 models
Video · KuaishouKling O3 (Omni)

The omni version of Kling 3.0, built to keep a character or product identical across clips. Pick it when the same face must come back shot after shot.

  • Latest : Kling 3.0 Omni (O3), February 2026
  • #14 of 22 on Artificial Analysis Image to Video Arena (1,044 Elo)
Anonymized Low cost9 models
Video · KuaishouKling 2.5 and 2.6

The previous Kling generation, kept because it is cheap and dependable. Sub Rosa's Studio workflows use 2.5 Turbo Pro by default.

  • Latest : Kling 2.6, December 2025
  • #19 of 22 on Artificial Analysis Image to Video Arena (988 Elo)
Anonymized Low cost4 models
Video · ByteDanceSeedance 2.5

ByteDance's July 2026 model, which makes up to 30 s in one pass with sound. Pick it for a long continuous scene.

  • Latest : Seedance 2.5, July 31, 2026
  • #2 of 18 on Artificial Analysis Text to Video Arena with audio (1,142 Elo)
Anonymized, some versions private Moderate cost6 models
Video · ByteDanceSeedance 2.0 and 1.5

ByteDance's 2026 multimodal model: image, references and audio in, a clip with synced sound out. Pick it for dialogue scenes up to 15 s.

  • Latest : Seedance 2.0, February 12, 2026
  • #18 of 22 on Artificial Analysis Image to Video Arena (1,000 Elo)
Anonymized Moderate cost11 models
Video · Google DeepMindVeo 3 and 3.1

Google's video model with dialogue, effects and ambience generated in the same pass. Pick it for short, polished, realistic shots.

  • Latest : Veo 3.1, October 15, 2025
  • #7 of 22 on Artificial Analysis Image to Video Arena (1,082 Elo)
Anonymized Low cost8 models
Video · AlibabaWan 3.0

Alibaba's August 2026 model, up to 30 s per clip with sound. Pick it for long shots at a mid-range price.

  • Latest : Wan 3.0, August 24, 2026
  • #3 of 22 on Artificial Analysis Image to Video Arena (1,164 Elo)
Anonymized Moderate cost12 models
Video · AlibabaWan 2.x

Alibaba's earlier Wan line (2.2 to 2.7), cheaper than Wan 3.0 and with an editing mode. Pick it for 5 to 15 s clips on a budget.

  • Latest : Wan 2.6, December 2025
  • #8 of 22 on Artificial Analysis Image to Video Arena (1,077 Elo)
Anonymized Low cost12 models
Video · MiniMaxMiniMax H3 (Hailuo)

MiniMax's 2026 Hailuo models with stereo sound. H3 Max runs in private mode at a low price. Pick it for private, affordable clips.

  • Latest : MiniMax H3 Max Recast, October 2026
  • #1 of 22 on Artificial Analysis Image to Video Arena (1,195 Elo)
Private, some versions anonymized Low cost12 models
Video · xAIGrok Imagine video

xAI's video model, every version running in private mode, with sound and any length from 1 to 15 s. Pick it to animate an image privately.

  • Latest : Grok Imagine video 1.5, May 30, 2026
  • #5 of 22 on Artificial Analysis Image to Video Arena (1,098 Elo)
Private Low cost9 models
Video · Google DeepMindGemini Omni Flash

Google's 2026 model that makes and edits video from text, images, references or footage. Pick it for short clips you want to refine by editing.

  • Latest : Gemini Omni 1.1 Flash, August 27, 2026
  • #7 of 18 on Artificial Analysis Text to Video Arena with audio (1,113 Elo)
Anonymized Low cost7 models
Video · AlibabaHappyHorse

Alibaba's 2026 model that generates picture and sound together, with lip-sync in seven languages including French. Pick it for talking characters.

  • Latest : HappyHorse 1.1, June 28, 2026
  • #4 of 22 on Artificial Analysis Image to Video Arena (1,104 Elo)
Anonymized Moderate cost7 models
Video · PixVerse (AIsphere)PixVerse

PixVerse C1 is an April 2026 model aimed at action and effects shots, with sound, at a low price per second. Pick it for cheap clips with sound.

  • Latest : PixVerse C1, April 7, 2026
  • #20 of 22 on Artificial Analysis Image to Video Arena (963 Elo)
Anonymized Low cost7 models
Video · RunwayRunway Gen-4

Runway's Gen-4.5 for realistic, controlled camera work from text or an image. Pick it for silent shots you will score or narrate yourself.

  • Latest : Runway Gen-4.5, December 1, 2025
Anonymized Low cost3 models
Video · LightricksLTX Video

Lightricks' open-weight models that make picture and sound together, up to 20 s and up to 4K. Pick them for long, sharp clips at a fair price.

  • Latest : LTX-2.5, August 2026
  • #15 of 22 on Artificial Analysis Image to Video Arena (1,038 Elo)
Anonymized Low cost8 models
Video · ShengShuVidu Q3

ShengShu's January 2026 model for short stories: 16 s, several camera angles and sound in one pass. Pick it for a small scene with dialogue.

  • Latest : Vidu Q3, January 30, 2026
  • #11 of 22 on Artificial Analysis Image to Video Arena (1,056 Elo)
Anonymized Moderate cost2 models
Video · MeituanLongCat Video

Meituan's open-source model built for long clips, running in private mode. Pick it for the lowest price per second or a 30 s silent shot.

  • Latest : LongCat-Video, October 27, 2025
Private Low cost4 models
Video · Character.AIOvi

An open-source model that animates an image into a 5 s clip with speech and sound, in private mode. Pick it for a cheap, private talking shot.

  • Latest : Ovi, October 5, 2025
Private Low cost1 model
Video · Black Forest LabsFLUX 3 video

Black Forest Labs' first video model (August 2026), with sound and clips up to 20 s. Pick it to link two images, to extend a clip or for a long shot with sound.

  • Latest : FLUX 3 video, August 4, 2026
  • #5 of 18 on Artificial Analysis Text to Video Arena with audio (1,127 Elo)
Anonymized Moderate cost4 models
Video · Topaz LabsTopaz Video Upscale

Not a generator: it enlarges and sharpens a clip you already have. Pick it to bring a finished 720p shot up to a higher resolution.

Anonymized1 model

Voice

Voice · hexgrad (open source)Kokoro

A small open-source voice model, private and by far the cheapest. Sub Rosa uses it by default for narration.

  • Latest : Kokoro v1.0, January 27, 2025
  • #10 of 10 on Artificial Analysis Speech Arena (1,064 Elo)
Private Low cost1 model
Voice · Resemble AIChatterbox HD

Resemble AI's expressive voice model, here in private mode. Pick it for expressive narration that stays private.

  • Latest : Chatterbox HD, May 2025
  • #7 of 10 on Artificial Analysis Speech Arena (1,102 Elo)
Private High cost1 model
Voice · ElevenLabsElevenLabs

ElevenLabs' voice models, from the fast Turbo v2.5 to the expressive Eleven v4. Pick Eleven v4 for natural narration in French or another of its 90+ languages (about 12 credits for a 3,000-character page).

  • Latest : Eleven v4 Turbo, September 28, 2026
  • #1 of 10 on Artificial Analysis Speech Arena (1,330 Elo)
Anonymized High cost5 models
Voice · Google DeepMindGemini 3.1 Flash TTS

Google's April 2026 voice model, steered with tags like [whispers] written into the text. Pick it when delivery matters more than price.

  • Latest : Gemini 3.1 Flash TTS, April 15, 2026
  • #3 of 10 on Artificial Analysis Speech Arena (1,209 Elo)
Anonymized Very high cost1 model
Voice · GradiumGradium TTS

A voice model from Gradium built for accurate pronunciation and low latency. Pick it for tricky names and terms read clearly.

  • Latest : Gradium TTS, August 2026
  • #5 of 10 on Artificial Analysis Speech Arena (1,151 Elo)
Anonymized High cost1 model
Voice · Inworld AIInworld TTS-1.5 Max

Inworld's January 2026 voice model: expressive, 15 languages, and the second cheapest voice here. Pick it for good quality on a budget.

  • Latest : TTS-1.5 Max, January 21, 2026
Anonymized Moderate cost1 model
Voice · MiniMaxMiniMax Speech-02 HD

MiniMax's high-fidelity voice model for voice-overs and audiobooks, in 30+ languages. Pick it for long, steady reading with character voices.

  • Latest : Speech-02 HD, April 2, 2025
  • #6 of 10 on Artificial Analysis Speech Arena (1,114 Elo)
Anonymized Very high cost1 model
Voice · Canopy Labs (open source)Orpheus TTS

An open-source English voice model that can laugh, sigh or gasp on cue, running in private mode. Pick it for lively English dialogue.

  • Latest : Orpheus 3B 0.1 finetuned, March 2025
Private High cost1 model
Voice · Alibaba (Qwen)Qwen3 TTS

Alibaba's open-source voice models, in 10 languages and running in private mode. Pick them for private narration in French or Asian languages.

  • Latest : Qwen3 TTS, January 22, 2026
Private Very high cost2 models
Voice · xAIxAI TTS

xAI's voice model with 26 voices, 20 languages and inline tags for pauses, laughter or whispers. Pick it for expressive speech at a low price.

  • Latest : xAI TTS, April 2026
Anonymized Moderate cost1 model
Voice · ByteDanceSeed Audio 1.0

ByteDance's audio model: you describe the voice, the performance and the background in one prompt. Pick it for a voiced scene with ambience.

Anonymized Moderate cost1 model

Models and prices change often. Sub Rosa always shows what is available today and the exact price before anything is charged. The search index models behind Sub Rosa’s memory are left out: the app picks them, you never do. What “private” and “anonymized” mean