AI 模型目錄

AI performance capture. Upload a character image + your performance video to transfer your motion to the character.

Runway's current video-to-video model. Edit existing footage with text prompts; accepts 2-30 second inputs.

Async's legacy low-latency model with the broadest language coverage (15 languages including Arabic, Russian, Japanese, Hebrew, Armenian, Turkish, Hindi and Chinese); speed and stability controls, no text normalisation.

Async's latency-optimised streaming model for real-time apps and voice agents: English, Spanish, French, German, Italian and Portuguese, with built-in normalisation of dates, currencies, numbers and abbreviations.

Async's highest-quality English model for content production and audiobooks, with built-in text normalisation; twice the price of Flash.

Deepgram's original text-to-speech model, 12 English voices at half the cost of Aura 2. Fast and reliable.

Deepgram's latest text-to-speech model, 90+ natural voices across 8 languages (EN, FR, DE, ES, IT, NL, JA), sub-200ms latency. Greek mythology-themed voice names.

Blur the background behind the subject

Sarvam's stable text-to-speech model for Indian languages and English. Returns base64-encoded WAV audio. Production-ready with consistent quality.

Sarvam's latest text-to-speech model, 30+ voices across Indian languages and English. Natural prosody with cultural intonation. Currently in beta.

ByteDance Seed 1.8, multimodal agent model (text, image & video in → text), strong tool use + reasoning, 256K context. Native via BytePlus (Ark).

ByteDance Seed 2.0 Code (preview), coding-specialized agent (text & image in → text), strong tool use + reasoning. Native via BytePlus (Ark).

ByteDance Seed 2.0 Lite, efficient multimodal agent (text, image & video in → text), tool use + reasoning, 256K context. Native via BytePlus (Ark).

ByteDance Seed 2.0 Mini, fast lightweight multimodal agent (text, image & video in → text), tool use. Native via BytePlus (Ark).

ByteDance Seed 2.0 Pro, flagship multimodal agent (text, image & video in → text), strong tool use + reasoning, 256K context. Native via BytePlus (Ark).

ByteDance Seed 2.1 Turbo, fast multimodal reasoning agent (text & image in → text), strong tool use + thinking, 256K context. Native via BytePlus (Ark).

English TTS with emotion control and zero-shot voice cloning

Multilingual TTS supporting 23 languages with natural prosody

Specialized model for code generation, completion, and understanding.

10-second, 4K, 60fps video generation with realistic physics.

SOTA open-source image model. Native Chinese text support, 6B parameters.

Bring black-and-white photos to natural color

Remove the background and crop tightly around the subject

High-quality neural machine translation supporting 33 languages

DeepSeek flagship V4. 512K context, advanced reasoning + tool use.

DeepSeek V4.1 Flash, the fast daily driver that replaced V4 Flash on 2026-09-10. 1M context, thinking (default) and non-thinking modes.

Most expressive TTS model with audio tags, dialogue mode, accent emulation, and 70+ languages. Best for long-form content.

Low-latency v3 tuned for realtime conversation (~280 ms), high-quality expressive delivery with custom audio tags, 70+ languages.

ElevenLabs' most expressive model (September 2026): tone, pacing and emotion read from the text, speaker identity held across long-form, audio tags, 85 languages, up to 10,000 characters per request.

Low-latency Eleven v4 for realtime voice (~100 ms median inference), audio tags, 85 languages, up to 10,000 characters per request.

Enhance image quality with AI

Remove the main subject and fill in the background

Paint over an object to remove it from the image

Expand image beyond its borders with AI outpainting

Fable 5 by Anthropic.

Fable 5.1 by Anthropic: the most capable generally available Claude, for demanding reasoning and long-horizon agentic work. 1M context, 128K output, adaptive thinking always on.

Ultra-low latency TTS for real-time and conversational AI. ~75ms latency.

The best of FLUX, offering state-of-the-art performance image generation at blazing fast speeds

Ultra / Raw Modes. Ultra Fast, More Pixels

FLUX 3 image generation: text-to-image, multi-reference editing with up to ten references, and bounding-box placement.

FLUX 3 multimodal video: text-to-video, image-to-video with up to 10 keyframes, and video continuation — up to 20 s in HD or FHD with native synchronized audio. Draft mode gives a fast, cheaper preview.

A premium model brings maximum performance across all aspects – greatly improved quality, consistency, and speed

A unified model delivering local editing, generative modifications, and text-to-image capabilities

Directly distilled from FLUX.1 [pro], an open-weight, guidance-distilled model for non-commercial use

Specialized for typography and text rendering with adjustable guidance and generation steps

Ultra-fast open-source FLUX model optimized for real-time generation at the lowest cost. Apache 2.0 licensed.

Fastest FLUX model with sub-second inference. 9B params with Qwen3 text embedder for superior prompt understanding.

Highest quality FLUX 2.0 with grounding search capability for photorealistic images up to 4MP

Production-grade FLUX 2.0 with multi-reference image editing and precise control over colors, poses, and composition

Sakana's multi-agent conductor, orchestrates a pool of frontier models for complex reasoning. Billed per token, orchestration included.

Fast and efficient model with adaptive thinking for complex tasks

Google's low-latency text-to-speech with natural prosody and controllable style. Supports 24 languages.

Cost-efficient high-throughput model for budget-conscious applications

Enhanced thinking and reasoning for complex problems

Premium text-to-speech with enhanced expressivity, richer tone, and precision pacing. Best for high-quality output.

Google's fastest frontier model. Beats 2.5 Pro at 1/4 the cost with 1M context

Latest price-performant, low-latency controllable speech generation. 30 voices, 24 languages.

Cost-efficient high-throughput model for budget-conscious applications.

Google's most capable agentic model with 1M token context, 77.1% ARC-AGI-2 reasoning, and native tool use.

Fast flagship with adaptive thinking for complex tasks.

Cost-efficient GA Flash model for high-throughput multimodal tasks.

Google's latest GA Flash model for fast multimodal reasoning and agentic workloads.

Google's most capable Flash model for agentic workflows and multimodal reasoning.

High-throughput, low-latency conversational speech generation. 30 curated voices, 24 languages.

Studio-grade speech generation with expressive acting and long-form stability. 30 prebuilt voices, 24 languages.

Cost-efficient text embeddings — 3072-dim vectors for search, clustering, and RAG.

Google's latest embedding model — 3072-dim vectors for retrieval and semantic search (multimodal upstream; text wired).

Google's open-weight mixture-of-experts model, 26B total, 4B active parameters for fast inference

Google's open-weight dense model with 262K context and strong multilingual reasoning

Transform reference images with text prompts. Preserves identity while changing pose, lighting, background.

Fast reference-based image generation. 2.5x faster, preserves identity and style.

Runway's flagship image-to-video model. Exceptional fidelity, stability, and controllability. Requires input image.

Runway's most capable video model. #1 on Artificial Analysis text-to-video benchmark. Text-to-video or image-to-video with exceptional physics and human motion.

Paint a mask and fill the area with AI-generated content

Previous flagship with thinking mode. MoE architecture, 128K context.

Lightweight model optimized for efficiency. 128K context.

Free tier model. Great for testing and light workloads.

Vision-language model for image understanding and analysis.

Flagship model with reasoning, coding, and agentic capabilities. 128K context.

Vision-capable model for image understanding and analysis.

Free vision model. Great for testing image understanding.

Lightweight vision model. Fast and cost-effective for image tasks.

Latest flagship. 358B params, 204K context, 131K output. #1 on LiveCodeBench.

Lightweight, completely free GLM-4.7 variant with 200K context.

Lightweight, high-speed and affordable GLM-4.7 variant with 200K context.

Most capable Z.ai model. 744B params (40B active MoE), 28.5T training tokens. Built for complex systems engineering and agentic tasks.

Fast variant of GLM-5 with optimized speed and competitive quality.

Z.ai's current flagship for agentic and coding tasks.

Z.ai's latest flagship for agentic and coding tasks.

Z.ai's newest flagship for agentic and coding tasks. Always reasons; text-only input.

Z.ai's fast, low-cost GLM-5.3 variant: 1M context, tools, always reasons; text and image input.

Native multimodal coding model. 744B MoE (40B active), 203K context. Optimized for design-to-code, GUI automation, and vision-grounded agentic tasks.

Z.ai's image generation model with strong prompt adherence and text rendering.

Specialized OCR model for extracting text from images and documents.

OpenAI's latest image generation model (high-quality tier).

OpenAI's ChatGPT Images 2.5, fast variant: everyday generation at about half the latency of GPT Image 2.

OpenAI's ChatGPT Images 2.5, precision variant: detailed creative work and editing fidelity, slower than Flare.

OpenAI GPT-3.5 Turbo (chat completions endpoint).

OpenAI GPT-3.5 Turbo 16K (chat completions endpoint).

OpenAI GPT-4 (chat completions endpoint).

OpenAI GPT-4.1 (chat completions endpoint).

OpenAI GPT-4.1 Mini (chat completions endpoint).

OpenAI GPT-4o (chat completions endpoint).

OpenAI GPT-4o Mini (chat completions endpoint).

OpenAI GPT-5 (chat completions endpoint).

OpenAI GPT-5 Mini (chat completions endpoint).

OpenAI GPT-5 Nano (chat completions endpoint).

OpenAI GPT-5 Pro (chat completions endpoint).

OpenAI GPT-5 Search API (chat completions endpoint).

OpenAI GPT-5.1 (chat completions endpoint).

OpenAI GPT-5.2 (chat completions endpoint).

OpenAI GPT-5.2 Pro (chat completions endpoint).

OpenAI GPT-5.3 Codex (chat completions endpoint).

OpenAI mid-tier flagship (March 2026).

OpenAI fast/cheap chat model.

OpenAI cheapest tier, very fast.

OpenAI GPT-5.4 Pro (chat completions endpoint).

OpenAI flagship chat model (April 2026 release).

OpenAI premium tier with deep reasoning. April 2026.

OpenAI GPT-5.6 Luna — high-volume streaming tasks, fast + economical 5.6 tier.

OpenAI GPT-5.6 Sol — frontier logic + deep reasoning, the flagship 5.6 tier.

OpenAI GPT-5.6 Terra — balanced business logic, mid 5.6 tier.

OpenAI GPT-6 Astra — frontier reasoning over a 1M-token context window.

OpenAI GPT-6 Luna: the fast, low-cost GPT-6 tier for high-volume chat and light agentic work, 1M-token context window.

OpenAI GPT-6 Sol: GPT-6 reasoning at a fifth of Astra's price, over a 1M-token context window.

Non-reasoning variant of Grok 4.20. Fastest 4.20 path for chat and high-throughput tasks.

Reasoning-mode variant of Grok 4.20. Slower but stronger on multi-step problems.

xAI's flagship general-purpose model. 1M context, balanced reasoning and chat.

xAI's newest flagship, launched July 2026. 500K context, strongest reasoning + coding, trained alongside Cursor.

xAI's newest flagship. 500K context, accepts text and images, strongest reasoning + coding.

xAI's newest flagship (2026-09-21). 500K context, accepts text and images, function calling, structured outputs, reasoning.

Compact model tuned for code generation and structured output. 256K context.

Fast text-to-image generation from xAI. Standard quality, lowest cost.

xAI's current image model. Higher fidelity than Grok Imagine, with stronger prompt adherence.

Higher-fidelity Grok Imagine. Better prompt adherence and detail, 2.5x the cost.

Text-to-video generation from xAI. ~8s clips, async render.

xAI Grok Imagine 1.5 — image-to-video: animates a starting frame from your prompt. Up to 1080p, 1-15s clips, async render.

xAI's expressive text-to-speech — 26 multilingual voices with inline speech tags for tone, pauses, whispers and laughter. 20+ languages with auto-detection.

Claude's fastest and most intelligent Haiku model

Haiku 5.5 by Anthropic (2026-10-07): the fastest Claude, for high-volume, latency-sensitive work such as classification, extraction and routing. 1M context, 128K output, adaptive thinking.

Higher resolution (1080p), longer duration (10s), stronger prompt adherence

Breakthroughs in body movement, facial expressions, and physical realism

Image-to-video model optimized for value and efficiency (requires image upload)

Arabic-first speech-to-text with dialect-aware recognition and broad multilingual coverage.

Arabic-first text-to-speech, low-latency tier, with 50 voices spanning Arabic dialects and 20+ languages.

Alibaba HappyHorse image-to-video. Animates a still image into a coherent motion sequence guided by text.

Alibaba HappyHorse reference-to-video. Generates video conditioned on a reference image plus text prompt.

Alibaba HappyHorse text-to-video. Generates short cinematic clips from text prompts.

Alibaba HappyHorse 1.1 image-to-video. Animates a still image into coherent motion guided by text (native DashScope, US-Virginia).

Alibaba HappyHorse 1.1 reference-to-video. Generates video that preserves the subject and scene from one or more reference images (native DashScope, US-Virginia).

Alibaba HappyHorse 1.1 text-to-video. Generates short cinematic clips from text prompts (native DashScope, US-Virginia).

Alibaba HappyHorse video editing. Edits a source video from natural-language instructions, optionally guided by reference images (native DashScope, US-Virginia).

Previous generation model. Still excellent for text rendering at lower cost.

State-of-the-art image generation with exceptional text rendering. Best balance of quality and speed.

Highest quality Ideogram 3.0. Maximum detail and fidelity for final outputs.

Faster Ideogram 3.0 variant. Quick generation for iterative workflows.

Ideogram 4.0 text-to-image generation with the best text rendering. The rendering speed is the priced tier: Turbo, Default or Quality.

Precise image editing on Ideogram 4.5: the first uploaded image is edited from the prompt (a mask marks what to change), up to four reference images steer the result, and unedited pixels are kept from the original.

Cinematic-quality text-to-image generation with high prompt fidelity

Jina AI's CLIP model, cross-modal text-image embeddings supporting 89 languages with Matryoshka representations. 865M params.

Jina AI's code-specialized embedding model, 1.5B parameters optimized for code search and retrieval. 32K context.

Jina AI's ColBERT late-interaction model, multi-vector embeddings for fine-grained retrieval, 89 languages, user-controlled embedding sizes.

Jina AI's 570M parameter multilingual text embedding model, MTEB benchmark champion, 8K context, 89 languages.

Jina AI's latest multimodal embedding model, embeds text and images into a shared vector space with 32K context. State-of-the-art retrieval quality.

Jina AI's multimodal reranker, re-ranks results containing both text and images for cross-modal search pipelines. On Zubnet it ranks text documents; image documents are not supported yet.

Jina AI's latest reranker, novel listwise architecture, SOTA multilingual retrieval, massive 131K context window. 0.6B params.

Latest Kimi K2 reasoning model. Replaces K2 thinking variants (k2-thinking and k2-thinking-turbo) being retired 2026-05-25.

Coding-specialized Kimi K2 model (k2.7). 256K context, agentic tool use and function calling, tuned for software engineering.

High-speed variant of Kimi K2.7 Code for low-latency agentic software engineering.

Moonshot's open-weight 2.8T-parameter multimodal reasoning model with a 1M-token context window.

High-quality text-to-image generation with multiple aspect ratios

Latest turbo model with fast text-to-video and image-to-video generation

First model with simultaneous audio-visual generation. Creates video with native speech, dialogue, sound effects, and ambient audio in Chinese/English. No post-production dubbing needed.

Cost-effective audio-visual generation with native speech, dialogue, and sound effects. Standard tier for balanced quality and cost.

Kling 3 image generation, text-to-image and image-to-image.

Kling 3 Omni unified image model, text-to-image, image-to-image, and image editing.

Kling 3 Omni, audio-visual generation with native sound, multi-shot, start/end frame & reference video. Standard tier.

Kling 3.0 Turbo, 720p, native audio included. Text-to-video and image-to-video, 3 to 15 seconds.

Kling 3.0 Turbo Pro, 1080p, native audio included. Text-to-video and image-to-video, 3 to 15 seconds.

Unified image model for text-to-image, image-to-image, and detail editing. Accepts up to 10 reference images for guided generation. Part of the O1 multimodal family.

Unified multimodal video model combining generation, editing, and transformation. Supports text, image, and video inputs with natural language control. Cost-effective tier.

High-speed anime-focused model excelling at a range of anime styles.

Core Leonardo model with stunning outputs and broad style range.

Cinematic-focused model excelling at film-like compositions and lighting.

High-speed generalist image generation model for fast iteration.

Versatile model excelling at realism and photography.

Generate sound effects for longer videos (SFX 1.6), up to 60 seconds of audio from a video URL.

Leonardo model for vibrant, diverse imagery in HD output.

Leonardo high-speed model designed for photorealistic outputs.

Google DeepMind's latest music generation — 30-second compositions from text prompts with SynthID watermarking.

Full-song generation from Google DeepMind's Lyria 3 — complete compositions from text prompts with SynthID watermarking.

Google DeepMind's Lyria 3.5 — full-length songs with verses, choruses and bridges, vocals and timed lyrics, 44.1 kHz stereo, SynthID watermarking.

Agentic capabilities with advanced reasoning and thinking process visibility

Optimized for high concurrency and commercial use with advanced reasoning

Frontier reasoning model with SOTA coding, agentic tool use, and complex real-world task performance

Fast variant of M2.5 optimized for speed at ~100 tokens per second

Autonomous real-world productivity with agentic collaboration, live debugging, and professional document generation

Fast variant of M2.7 for latency-sensitive work. 200K context, text only.

MiniMax flagship LLM, 1M context.

Efficient reasoning model for everyday logic tasks.

Transform images into detailed 3D models with textures and PBR maps

Generate detailed 3D models from text descriptions with PBR textures

Transform images into detailed 3D models with the highest image alignment Meshy has shipped, textures and PBR maps

Combine two to four views of the same subject into one aligned, textured 3D model with PBR maps

Transform images into detailed 3D models with Meshy's sharpest geometry yet, textures and PBR maps

Combine two to four views of the same subject into one aligned, textured 3D model with Meshy's sharpest geometry yet

Generate detailed 3D models from text descriptions with Meshy's sharpest geometry yet, textured with PBR maps

Xiaomi's fast, low-cost MiMo model (V2.6, 2026-09). 1M context, accepts text and images, function calling and structured output, reasons before it answers.

Xiaomi's flagship MiMo model (V2.6, 2026-09). 1M context, accepts text and images, function calling and structured output, reasons before it answers.

MiMo V2.6 Pro on Xiaomi's high-speed tier, at a premium price. 1M context, accepts text and images, function calling and structured output.

High-performance edge model with vision. Best Ministral for complex tasks.

Ultra-lightweight edge model with vision. 3B params, runs on phones/laptops.

Compact edge model with vision. 8B params for local deployment.

Mistral's text embedding model for semantic search, retrieval, and classification.

Open-weight flagship. 675B MoE (41B active), multimodal with vision. Apache 2.0 licensed.

Premier frontier-class multimodal model. Great balance between Large and Small.

Mistral's content moderation model for detecting harmful, unsafe, or policy-violating content.

Document AI model for high-accuracy OCR. Extracts text from images, PDFs, handwritten content, forms, and complex tables.

Fast and efficient model for simple tasks. Great balance of performance and cost.

Mistral's model optimized for command-line and terminal-based coding workflows.

Most advanced emotionally-aware speech synthesis with rich expression across 29 languages

ElevenLabs' value-tier music generation: songs up to 10 minutes from a prompt, instrumental on request. Billed per generated minute upstream.

ElevenLabs' studio-grade music generation: songs up to 10 minutes from a prompt, instrumental on request. Billed per generated minute upstream.

Gemini 2.5 Flash Image, optimized for speed and efficiency, high-volume low-latency image generation.

High-efficiency image generation, optimized for speed and high-volume use.

Efficiency specialist of the Nano Banana family — ultra-low-latency, cost-effective image generation and editing.

Google's native image generation with text rendering and multimodal understanding.

OpenAI o3 (chat completions endpoint).

Premium model combining maximum intelligence with practical performance

Most capable Claude model with enhanced reasoning and coding

Highly capable Claude Opus for complex reasoning and coding (1M context)

Anthropic's most capable model for complex reasoning and agentic coding (1M context)

Anthropic's flagship Opus — deep reasoning, agentic coding and long-horizon work (1M context)

Opus 5.5 by Anthropic (2026-09-21): the next Opus, for long-running agentic coding and knowledge work. 1M context, 128K output, adaptive thinking always on.

Large-context generalist with 128K window. Excellent for complex generation and analysis. Augure sovereign Canadian AI.

Premium multimodal model with vision and tool use, strong bilingual EN/FR. Augure sovereign Canadian AI.

Newest reasoning model: the fastest of the Augure menu, strong at code, 256K context. Augure sovereign Canadian AI.

Leonardo foundational model preview with extreme prompt adherence and versatility.

Leonardo flagship foundational model with exceptional prompt adherence and text rendering.

PixVerse C1, cinematic-quality video generation tuned for action scenes and reference-based work, clips from 1 to 15 seconds with optional generated audio.

PixVerse C1 image-to-video: cinematic animation of a source image into a 1–15 second clip with optional generated audio.

PixVerse C1 fusion: cinematic video from 1–3 reference images (subjects or scenes) guided by a prompt, with optional generated audio.

PixVerse V6, general-purpose video generation with clips from 1 to 15 seconds, optional generated audio, and multi-clip support.

PixVerse V6 image-to-video: animate a source image into a 1–15 second clip with optional generated audio.

PixVerse V6 fusion: generate a video from 1–3 reference images (subjects or scenes) guided by a prompt, with optional generated audio.

Perplexity's lightweight embedding model. 1024-dimensional INT8 vectors, 32K context, Matryoshka dimension reduction.

Perplexity's most capable embedding model. 2560-dimensional INT8 vectors, 32K context, Matryoshka dimension reduction.

Pruna's optimized text-to-image: fast, high-quality generation.

Pruna's optimized image editing: transform an input image with a prompt.

Pruna's premium video generation: text-to-video and image-to-video, up to 1080p. Draft mode gives fast, cheap previews.

Ultra-fast model with 1M context. Best latency and cost efficiency for simple tasks.

Alibaba's flagship text-to-image model — successor to the whole Qwen-Image family, with strong prompt following and text rendering (native DashScope, Singapore).

The higher-fidelity "pro" tier of Alibaba's Qwen Image 2.0 — stronger prompt following and text rendering (native DashScope, Singapore).

Alibaba's Qwen Image 3.0 — the next generation of the Qwen text-to-image family: denser layouts, images within images, sharper small text, native rendering in 12 languages (native DashScope, Singapore).

The higher-fidelity pro tier of Alibaba's Qwen Image 3.0 — prompts up to 4.5k tokens, dense information layouts, precise text down to 10 px (native DashScope, Singapore).

Balanced flagship model with 1M context. Best performance/cost ratio for most tasks.

Cost-effective coding model optimized for speed. Fast code generation and completion at lower cost.

Advanced vision-language model with 262K context. Excels at visual coding, spatial perception, and multimodal reasoning.

Flagship 1T+ parameter model. State-of-the-art reasoning, coding, and agent capabilities.

Alibaba Qwen3.5 Omni Flash, fast multimodal (text & image input, text output). Singapore deployment; native voice/audio coming.

Alibaba Qwen3.5 Omni, multimodal (text & image input, text output). Singapore deployment; native voice/audio coming.

397B MoE model with 17B active parameters, 1M token context, multimodal (text/image/video)

Flagship 1T+ parameter model. State-of-the-art reasoning, coding, and agent capabilities.

Qwen3.6 Plus, next-gen frontier LLM (Agentic Coding, vision/OCR, fine-grained localization). Replaces Qwen3.5 Plus.

Alibaba's Agent Frontier flagship.

Multimodal frontier (vision/video); cheaper than Max.

Alibaba's 2.4T-parameter multimodal MoE flagship. 1M context, accepts text and images.

Fast multimodal Qwen (text & image in, text out) on the Singapore tenant. Thinking on by default. Context measured served: 1,006,592 tokens.

Advanced AI video generation model by Luma AI

Advanced AI video generation model by Luma AI

Advanced AI video generation model by Luma AI

Advanced AI video generation model by Luma AI

Advanced AI video generation model by Luma AI

Advanced AI video generation model by Luma AI

Classic pixel art generation with 15 styles. Great for game assets, characters, textures, and retro-styled visuals. 16-512px.

Advanced pixel art with Flux model. Same 15 styles as Classic with enhanced quality and detail. 16-512px.

Recraft V3 (Red Panda), supports text-to-image, image-to-image, inpainting, and background replacement. 1024x1024.

Recraft V3 for vector graphics, SVG output with image-to-image, inpainting, and background replacement support.

Recraft's latest image generation model, photorealistic and illustration styles, text rendering, 1024x1024, ~10s generation.

Recraft's highest quality image model, 2048x2048 output, superior anatomy and detail, ideal for print-ready assets. ~30s generation.

Recraft's highest quality vector model, high-resolution SVG output with fine detail, ideal for logos and brand assets.

Recraft V4 for scalable vector graphics, generates production-quality SVG with discrete color regions and clean geometry.

Remove image background with transparency

Replace image background with AI-generated scene

Voyage AI's latest reranking model, re-scores search results for maximum relevance. Best quality. 200M free tokens.

Voyage AI's fast reranking model, cost-efficient re-scoring for high-volume retrieval pipelines. 200M free tokens.

Voyage AI's next-generation reranker, strongest on long documents and code. 32K-token context.

Voyage AI's fast next-generation reranker, cost-efficient re-scoring for high-volume retrieval. 32K-token context.

Repair and revive old or damaged photos

Rime's flagship voice model, 269 voices across 9 languages with rich emotional range.

Stylized voices in 4 categories (Professional / Formal / Casual / Energetic). 184 voices, 8 languages.

Original Mist model, 117 English voices, fast.

Mist generation 2, 141 voices across 4 languages.

Latest Mist generation, 83 voices, 4 languages. Improved expressiveness.

Premium model for complex, multi-step work. Augure sovereign Canadian AI.

Sarvam's 24B multilingual LLM, fluent in all 22 official Indian languages + English. OpenAI-compatible API. Free tier with no per-token charges. Supports wiki grounding, reasoning effort control, and tool calls.

Batch speech recognition with word-level timestamps and language detection across 99 languages.

Batch state-of-the-art speech recognition: 98%+ accuracy, keyterm prompting, 90+ languages.

Speech recognition fine-tuned for clinical audio: 35% fewer clinical errors than Scribe v2.

Ultra-low latency (<150ms) live speech recognition. 93.5% accuracy across 90+ languages. WebSocket streaming with VAD.

Balanced speed and quality for high-volume digital asset creation

Highest-quality photo-realistic image generation perfect for professional print media

8B parameter MMDiT model. Superior quality, typography, and prompt adherence. Most powerful in SD family.

Distilled SD3.5 Large optimized for faster inference with fewer steps.

Balanced SD3.5 model. Great quality with lower resource requirements than Large.

SDXL 1.0 for high-resolution 1024x1024 image generation

ByteDance Seedance 1.5 Pro, text & image to video with native synced audio (voice, SFX, music), via BytePlus (Ark). Billed per output token; audio videos bill a higher rate.

ByteDance Seedance 2.0, text & image to video, native via BytePlus (Ark). Billed per output token.

ByteDance Seedance 2.0 Fast — speed-optimized text & image to video, native via BytePlus (Ark). Billed per output token.

ByteDance Seedance 2.0 Mini — cost-optimized text & image to video, native via BytePlus (Ark). Billed per output token.

ByteDance Seedance 2.5, text & image to video, native via BytePlus (Ark). Billed per output token.

ByteDance Seedream 5.0 Lite, text-to-image (2K+), native via BytePlus (Ark).

ByteDance Seedream 5.0 Pro, flagship text-to-image (2K+), highest quality + consistency. Native via BytePlus (Ark).

Speechify's streaming-native multilingual model: English plus German, Spanish (Spain and Mexico), French, Italian and Brazilian Portuguese, and it accepts every Speechify voice whatever its locale.

Speechify's recommended streaming-native English model: lowest time-to-first-byte in the Simba family, fine-grained emotional control and SSML prosody. English voices only.

Upstage compact agentic MoE (35B total, 3B active). 524,288-token context (API-confirmed), tool calling and opt-in reasoning, English/Korean/Japanese.

Upstage nightly build of Solar Pro 2 with latest improvements.

Upstage flagship agentic model. 512K context (API-confirmed family), tool calling, structured outputs and opt-in reasoning. English/Korean/Japanese.

Perplexity's lightweight search-augmented model, real-time web search with citations, fast responses, 128K context.

Perplexity's most thorough research model, multi-step deep web investigation with comprehensive citations and reasoning.

Perplexity's advanced search model, deeper web research, multi-step reasoning with citations, 200K context.

Perplexity's reasoning model, extended thinking with real-time web search, ideal for complex research and analysis.

Cartesia's most natural streaming text-to-speech (2026-08-27): pacing and intonation from context, 60+ emotions, speed and volume controls, 44 languages including Odia and Urdu, sub-100 ms first audio.

Claude's best model for complex agents and coding

Best Claude Sonnet for complex agents and coding (1M context)

Anthropic's most capable Sonnet — closes the gap with Opus for complex agents and coding (1M context)

Sonnet 5.5 by Anthropic (2026-09-28): the best combination of speed and intelligence, for everyday coding, documents and agents. 1M context, 128K output, adaptive thinking.

Sound effects from a text description: automatic length or a set length up to 22 seconds, with adjustable prompt influence.

Focuses on rhythm, stability, and high-quality voice replication

Enhanced multilingual capabilities with turbo speed

Latest HD variant emphasizing prosody and voice cloning quality

Turbo performance with 40 language support and low latency

Current HD voice model. Highest fidelity prosody and voice cloning.

Current Turbo voice model. Low latency, broad language coverage.

Generate music and sound effects up to 3 minutes from text prompts. Produces structured compositions with intros, development, and outros at 44.1kHz stereo.

Enterprise-grade music and sound generation. Produces structured compositions with intros, development, and outros at 44.1kHz stereo. 8-step inference for fast generation.

Separate audio into individual stems (vocals, drums, bass, etc). 2-stem or 6-stem modes.

StepFun's latest reasoning model, 196B MoE (11B active), 256K context, open-source Apache 2.0. Fast agentic intelligence with strong code and math.

StepFun's multimodal flagship reasoning model.

Better song organization with clear verse/chorus patterns

Enhanced vocal quality and refined audio processing for music generation

Excellent prompt understanding with faster generation speeds, supports up to 8 minute tracks

Advanced model with enhanced tonal variation and excellent prompt understanding

Cutting-edge model with enhanced quality and capabilities for AI music generation

Upstage advanced synthetic reasoning model with 64K context.

Studio-grade lip sync. Syncs video lip movements to match any audio input with natural speaker style preservation.

Premium lip sync with enhanced detail preservation for beards, teeth, and fine facial features using diffusion-based super resolution.

Light tier: fast, thinking off, 256K context, served in Canada with no failover outside Canada. Replaces Tofino 2.5, retired by Augure on 2026-09-19.

Tripo's P1 model for image-to-3D, converts a single image into a game-ready low-poly mesh (50 to 20k faces) with clean topology and PBR textures.

Tripo's P1 model for multiview reconstruction, builds a game-ready low-poly mesh with clean topology from up to 4 view images (front required).

Tripo's P1 model for text-to-3D, generates game-ready low-poly meshes (50 to 20k faces) with clean topology and PBR textures. Exports GLB, FBX, OBJ, USD, STL.

Tripo's next-gen P2.0 Preview model for image-to-3D, converts a single image into a native quad mesh (48 to 25k four-sided faces) ready for editing and subdivision, with PBR textures.

Tripo's next-gen P2.0 Preview model for image-to-3D, converts a single image into a low-poly mesh (48 to 50k faces) with cleaner topology and PBR textures.

Tripo's next-gen P2.0 Preview model for multiview reconstruction, builds a native quad mesh (48 to 25k four-sided faces) ready for editing and subdivision from up to 4 view images (front required).

Tripo's next-gen P2.0 Preview model for multiview reconstruction, builds a low-poly mesh (48 to 50k faces) with cleaner topology from up to 4 view images (front required).

Tripo's next-gen P2.0 Preview model for text-to-3D, generates native quad meshes (48 to 25k four-sided faces) ready for editing and subdivision, with PBR textures. Exports GLB, FBX, OBJ, USD, STL.

Tripo's next-gen P2.0 Preview model for text-to-3D, generates low-poly meshes (48 to 50k faces) with cleaner topology and PBR textures. Exports GLB, FBX, OBJ, USD, STL.

Splits an existing 3D model into separately editable parts, textures and PBR materials preserved. A practical level of separation for most assets.

Splits an existing 3D model into many separately editable parts, textures and PBR materials preserved. Maximum component separation for deep editing and production.

Splits an existing 3D model into a few separately editable parts, textures and PBR materials preserved. Quick separation for review and lightweight prep.

Tripo's balanced v2.5 model for converting a single image into a 3D mesh, extracts geometry, texture, and materials from a photo or illustration.

Tripo v2.5 multiview generation, reconstructs detailed meshes from multiple image perspectives for improved accuracy.

Tripo's balanced v2.5 model for generating 3D meshes from text descriptions, produces detailed geometry with PBR materials in seconds. Exports GLB, FBX, OBJ, USD, STL.

Tripo's stable v3.0 model for image-to-3D, reconstructs clean geometry, textures, and PBR materials from a single image, up to 2M polygons in Ultra mode.

Tripo v3.0 multiview generation, reconstructs detailed meshes from up to 4 view images (front required) with stable, production-proven quality.

Tripo's stable v3.0 model for text-to-3D, crisp edges and coherent structure with PBR materials, up to 2M polygons in Ultra mode. Exports GLB, FBX, OBJ, USD, STL.

Tripo's flagship v3.1 model for image-to-3D, reconstructs high-fidelity geometry, textures, and PBR materials from a single image, up to 2M polygons in Ultra mode.

Tripo's flagship v3.1 image-to-3D with 8K Ultra textures, reconstructs maximum-fidelity geometry and materials from a single image for hero assets.

Tripo v3.1 image-to-3D in part-segmentation mode, converts a single image into an untextured mesh split into separately editable parts.

Tripo's highest-fidelity 3D generation, v3.1 reconstructs detailed meshes from up to 4 view images (front required) for maximum accuracy.

Tripo's highest-fidelity pipeline, v3.1 multiview reconstruction with 8K Ultra textures from up to 4 view images (front required).

Tripo v3.1 multiview reconstruction in part-segmentation mode, builds an untextured segmented mesh from up to 4 view images (front required).

Tripo's flagship v3.1 model for text-to-3D, sculpture-level geometry with crisp edges and PBR materials, up to 2M polygons in Ultra mode. Exports GLB, FBX, OBJ, USD, STL.

Tripo's flagship v3.1 text-to-3D with 8K Ultra textures, maximum-fidelity materials and fine surface detail for hero assets and close-ups. Exports GLB, FBX, OBJ, USD, STL.

Tripo v3.1 text-to-3D in part-segmentation mode, generates an untextured mesh split into separately editable parts for editing, rigging prep, and 3D printing. Exports GLB, FBX, OBJ, USD, STL.

Turns a single image into a photorealistic 3D Gaussian Splat in under a minute. Ideal for AR/VR, web scenes and visualization — a view-ready format, not an editable or printable mesh. Exports SPLAT.

Low-latency speech generation in 32 languages, optimized for real-time conversational AI

The most creative, intelligent and personalizable image generation models built on a new groundbreaking architecture that delivers ultra high quality and 10x higher cost efficiency.

The most creative, intelligent and personalizable image generation models built on a new groundbreaking architecture that delivers ultra high quality and 10x higher cost efficiency.

Double image resolution

Quadruple image resolution

Latest Veo with video extending capability. Create and extend AI-generated videos with improved consistency.

Fast variant of Veo 3.1. Quick video generation and extending with good quality.

Cost-efficient Veo 3.1 tier — video with audio at a fraction of the price (720p/1080p, no 4K).

Generate and edit sound effects from video (SFX 1.6). Provide a video URL and optional text prompt; adds seamless extension, looping ambiences, and AI inpainting to erase/replace moments.

Image-to-video generation. 4-second 720p output.

Reference-based video generation using multiple images for visual consistency. 4-second 720p output.

Start- and end-frame video generation. 5-second 720p output.

Create a talking avatar from an image or video with AI-generated speech from text

Synchronize video lips to an audio file for realistic dubbing and voice replacement

High-quality text-to-video with anime style support. 5-second 1080p output.

Image-to-video generation. 5-second 1080p output.

Start- and end-frame video generation. 5-second 1080p output.

Generate images from 1-7 reference images. Upload reference images to get started.

Generate images from text or reference images. Supports 1080p, 2K, and 4K resolution.

Ultra-fast video from text or image. Upload an image for image-to-video mode.

Natural text-to-speech with adjustable speed, volume, and pitch.

Transform a recording into another voice, keeping the words, timing and emotion. English only.

Transform a recording into another voice, keeping the words, timing and emotion. Multilingual.

Voyage AI's balanced general-purpose embedding model, strong quality at moderate cost. 200M free tokens.

Voyage AI's most capable general-purpose embedding model, highest quality retrieval and semantic search. 200M free tokens.

Voyage AI's fastest and cheapest embedding model, ideal for high-volume, low-latency use cases. 200M free tokens.

Voyage AI's code-optimized embedding model, best-in-class for code search, retrieval, and similarity. 200M free tokens.

Voyage AI's long-context embedding model, optimized for large documents and RAG with extended context windows. 200M free tokens.

Voyage AI's multimodal embedding model, embeds both text and images into a shared vector space for cross-modal search.

Alibaba Wan 2.2 Animate Mix, composite a character into a reference video (native DashScope, Singapore).

Alibaba Wan 2.2 Animate Move, animate a character image with a reference video's motion (native DashScope, Singapore).

Alibaba Wan 2.6 image editing. Edits 1-4 input images from a natural-language instruction (native DashScope, US-Virginia).

Alibaba Wan 2.6 image-to-video. Animates a still image into video while preserving subject, style and detail (native DashScope, US-Virginia).

Alibaba Wan 2.6 text-to-image. Photorealistic generation with accurate text rendering and flexible artistic styles (native DashScope, US-Virginia).

Alibaba Wan 2.6 reference-to-video. Generates video preserving the look (and voice) of subjects from a reference video and/or reference images (native DashScope, US-Virginia).

Alibaba Wan 2.6 text-to-video. Cinematic motion generation from text with strong instruction following (native DashScope, US-Virginia).

Alibaba Wan 2.7 image-to-video. Animates a still image into video with first-frame fidelity, clips up to 15 seconds (native DashScope, US-Virginia).

Alibaba Wan 2.7 text-to-image and editing. Served from Singapore — the US tenant returns AccessDenied.

Alibaba Wan 2.7 pro tier, up to 4K output for print and large-format work. Served from Singapore.

Alibaba Wan 2.7 text-to-video. First and last frame control, clips up to 15 seconds, sharper motion and instruction following (native DashScope, US-Virginia).