为下一个想法，找到合适的模型。

OpenAI GPT Image 2 generation and reference-image editing via OpenAI official, with Vercel AI Gateway and fal.ai fallbacks

Fast everyday image generation and reference editing. Uses the exact OpenAI model through Vercel AI Gateway; billed at reported upstream cost.

Precise image generation and reference editing. Uses the exact OpenAI model through Vercel AI Gateway; billed at reported upstream cost.

Fastest and cheapest Claude. Within ~5 percentage points of Sonnet 4.5 on many benchmarks at roughly one-third the cost.

Fast Claude model for high-volume agent, routing and tool-use workloads. Base token pricing for prompts up to 100,000 tokens.

Anthropic's most capable model. Use for long-context analysis, ambiguous research-style problems, and the hardest agentic tasks.

Anthropic's most capable model, for the hardest coding, agentic, creative and long-context work.

Anthropic's balanced flagship — strong reasoning, vision, and tool-use at a moderate price. Used by Claude Code by default.

Anthropic legacy Sonnet generation, superseded by Claude Sonnet 5; kept for workloads pinned to its behaviour.

Anthropic's most capable Sonnet-class model, built for coding, agents, professional work, and long-context multimodal tasks.

Codestral Latest routed by the Aiduct API gateway.

Strong reasoning at a very low price point. Competitive with GPT-4o-mini for code and structured-output workloads.

Deepseek R1 routed by the Aiduct API gateway.

Deepseek V3 routed by the Aiduct API gateway.

Deepseek V3 Siliconflow routed by the Aiduct API gateway.

DeepSeek's July 31 V4 Flash release for long-context reasoning, structured output, and agent tool use.

DeepSeek's August 13 V4 Pro release for long-context reasoning, structured output, and agent tool use.

DeepSeek's September 10 V4.1 Flash release with image understanding, long-context reasoning, JSON output, and agent tool use.

Generate a 60-second song or instrumental track from a text description. This platform model fixes the duration to 60 seconds and the output to MP3; each completed track is billed once. Returns an asynchronous audio polling ticket.

Development model with good quality-speed balance

High-quality professional image generation

Fastest Flux model for rapid iteration

Gemini 2.5 Flash routed by the Aiduct API gateway.

Gemini 2.5 Pro routed by the Aiduct API gateway.

Google's latest agentic Flash model for coding, tool use, long-context reasoning, and multimodal chat.

Google's most capable Flash model for agentic workflows, long-context multimodal reasoning, and tool use.

Google's newest Flash model for fast agentic workflows, long-context multimodal reasoning, and tool use.

Gemini native single-speaker text-to-speech with Gemini's prebuilt voices.

Gemma2 9b It routed by the Aiduct API gateway.

Glm 4 Air routed by the Aiduct API gateway.

Glm 4 Plus routed by the Aiduct API gateway.

GLM-5.2 routed by the Aiduct API gateway.

Z.ai's GLM-5.3 for long-context reasoning, structured output, and agent tool use.

Z.ai's fast GLM-5.3 variant for high-volume agent turns, with reasoning and tool use at a fraction of the flagship price.

Multilingual music with vocals, lyrics, or instrumental arrangements described in the input. MP3 output through the asynchronous audio queue.

High-quality multilingual songs and arrangements with vocals and lyrics described in the input. MP3 output through the asynchronous audio queue.

OpenAI's flagship multimodal model. Strong reasoning, tool use, and vision; the right default when GPT-4o-mini is not capable enough.

OpenAI's small, fast, cheap multimodal model. The default choice for routine chat, structured output, and short coding tasks.

OpenAI's GPT-6.1 Sol for coding, computer use and professional work, with adjustable reasoning effort.

Grok 4.5 routed by the Aiduct API gateway.

xAI's Grok 4.6 for agent tool use, long-context reasoning, and multimodal input.

Strong motion model for action-heavy scenes, fights, and kinetic camera work

Hunyuan Large routed by the Aiduct API gateway.

Hunyuan Turbos Latest routed by the Aiduct API gateway.

Tencent HunyuanVideo: open-source 5-second text and image-to-video generation.

Excellent at text rendering and typography in images

Kimi K2 0711 Preview routed by the Aiduct API gateway.

Kling 1.6 Pro cinematic video generation

Kling 1.6 Standard cost-efficient video generation

Kling 2.0 Master premium cinematic video generation

High-quality cinematic video generation

Kling 2.1 standard tier for lower-cost image-to-video generation.

Kling 2.5 Turbo Pro for fast text and image-to-video with first/last frames.

Kling 2.5 Turbo standard image-to-video oriented tier for fast motion drafts and short clips.

Kling 2.6 Pro text and image-to-video tier with cinematic motion and native-audio capable upstream support.

Kling 3.0 standard tier for cinematic text and image-to-video generation.

Kolors 1.0 text-to-image generation

Kolors 1.5 text-to-image generation

Kolors 2.0 high-fidelity text-to-image generation

Kling 3.0 Pro with cinematic motion, native audio, and 3-15 second clips

Llama 3.1 8b Instant routed by the Aiduct API gateway.

Llama 3.3 70b routed by the Aiduct API gateway.

Llama 3.3 70b Instruct routed by the Aiduct API gateway.

Llama 3.3 70b Versatile routed by the Aiduct API gateway.

Llama V3p1 405b Instruct routed by the Aiduct API gateway.

High-quality creative video with loop support

Fast variant of Ray2 for rapid video iteration

Meta Llama 3.1 405b Instruct Turbo routed by the Aiduct API gateway.

Meta Llama 3.1 70b Instruct Turbo routed by the Aiduct API gateway.

MiniMax's latest native multimodal video model with 2K output, first/last-frame control, and image, video, and audio references

fal's post-trained MiniMax H3 variant for stronger prompt adherence, cinematic aesthetics, and faster-than-realtime generation

High-quality AI music and sound effect generation. Async-only — POST returns a polling ticket like /v1/videos/generations because generation takes 30s-2min upstream.

MiniMax Speech-02 HD text-to-speech with MiniMax's multilingual system voices.

MiniMax Video-01: fixed 6-second 720p text or image-to-video with prompt optimization.

Mistral Large Latest routed by the Aiduct API gateway.

Mistral Nemo 12b Instruct routed by the Aiduct API gateway.

Mistral Small Latest routed by the Aiduct API gateway.

Mixtral 8x7b 32768 routed by the Aiduct API gateway.

Mixtral 8x7b Instruct routed by the Aiduct API gateway.

Moonshot V1 128k routed by the Aiduct API gateway.

Original Nano Banana for rapid iteration at low cost

Fast reasoning-guided image generation with text rendering and character consistency

Highest quality image generation with deep reasoning and up to 4K resolution

Reasoning model with adjustable effort. Used by Codex CLI for harder bugs and longer-horizon code tasks.

PixVerse v4 — previous generation, lower cost than v4.5.

PixVerse v4.5 — stylized video generation with strong anime/stylized motion.

Qwen 3 32b routed by the Aiduct API gateway.

Alibaba's 27B instruction-tuned Qwen 3.8, tuned for agentic tool use and long-context work.

Qwen Max routed by the Aiduct API gateway.

Qwen Plus routed by the Aiduct API gateway.

Qwen Turbo routed by the Aiduct API gateway.

Qwen2.5 72b Instruct Siliconflow routed by the Aiduct API gateway.

Qwen2.5 72b Instruct Turbo routed by the Aiduct API gateway.

Qwen2p5 72b Instruct routed by the Aiduct API gateway.

Qwq Plus routed by the Aiduct API gateway.

Professional design and illustration generation

Stable Diffusion 3.5 Large model

Official Seedance 1.0 quality model for text, first-frame, and first/last-frame video generation

Official lower-latency Seedance 1.0 variant for text and first-frame video generation

Official Seedance 1.5 Pro with native audio and first/last-frame control; scheduled for Volcengine EOS on 2026-09-21

Most capable general-purpose video model with multimodal image, video, and audio references

Lower-latency, lower-cost Seedance 2.0 variant with multimodal reference support

Lower-latency Seedance 2.0 multimodal reference-to-video workflow for image, video, and audio guided short-form generation.

Cost-efficient Seedance 2.0 variant with multimodal image, video, and audio references

Seedance 2.0 multimodal reference-to-video workflow for image, video, and audio guided character, motion, scene, and rhythm consistency.

ByteDance's latest professional multimodal video model with up to 30-second native-audio generation and stronger reference control

ByteDance Seedream 3.0 text-to-image model for fast creative drafts and general image generation.

ByteDance Seedream 4.0 text-to-image model with strong text rendering, composition, and commercial image workflows.

ByteDance Seedream 4.5 text-to-image model for high-resolution commercial image generation and prompt adherence.

Sonar routed by the Aiduct API gateway.

Sonar Pro routed by the Aiduct API gateway.

Sonar Reasoning routed by the Aiduct API gateway.

OpenAI's flagship video generation model. Strong physics, instruction following, and synchronized audio. Routed via fal.ai until the OpenAI provider gains a video adapter.

OpenAI's strongest embedding model. 3072-dim default; use when retrieval quality matters more than cost.

OpenAI's most cost-effective embedding model. 1536-dim default with `dimensions` truncation supported down to 256.

OpenAI's standard text-to-speech model. Fast, natural-sounding voices (alloy, echo, fable, onyx, nova, shimmer). Use for conversational latency-sensitive workloads.

Higher-fidelity OpenAI text-to-speech. Same voices as tts-1 with cleaner audio; pick this for narration and consumer-facing audio.

Google Veo 3.0 — previous generation, cheaper than Veo 3.1 for general clips.

Latest Google video generation model with highest quality

Google Veo 3.1 Fast: lower-latency Veo 3.1 with native audio, first/last frames and reference images.

High-efficiency Google Veo 3.1 model for lower-cost text-to-video and image-to-video generation with audio.

Natural text-to-speech using fal.ai's verified VibeVoice 7B presets

Vidu Q2 image-to-video route for animating a starting frame.

Vidu Q2 reference image-to-video route for subject consistency with multiple reference images.

Vidu Q2 text-to-video route with fast 2-8 second generation.

Q3 multi-reference image-to-video workflow for consistent subjects across scenes

Vidu Q3 image-to-video route for animating a starting frame.

Latest Vidu Q3 text and image-to-video path with audio-enabled 1-16 second clips

Vidu Q3 multi-reference image-to-video route for consistent subjects across scenes.

Vidu Q3 text-to-video route for prompt-only generation.

Legacy start-frame workflow for 5, 10, or 15 second clips with optional soundtrack input

Faster legacy Wan start-frame workflow for lower-cost iterations

Alibaba Wan 3.0 all-in-one video generation with text, first/last frames, image, video, and audio references, native audio, and clips up to 30 seconds

OpenAI speech-to-text transcription model exposed through the gateway /v1/audio/transcriptions primitive.