fal.ai
Listed at https://fal.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 按请求/按时长计费，无月费。图像生成：FLUX.2 Pro $0.03/第一个 MP + $0.015/额外 MP、FLUX1.1 [pro] 按 MP、Nano Banana 2 $0.08/图、FLUX.1 [schnell] 无定价显示。视频：Kling v3 $0.084-0.168/秒、Seedance 2.0 $0.3034/秒(720p)。TTS：通过 ElevenLabs/MiniMax 等第三方提供。余额模式：预付 $10+ 存入信用余额，用完为止。开源模型按秒计费，秒级颗粒度。 | per million tokens |
| Output | per million tokens |
🤖 Supported Models (1396)
✨ Pros
- ✓🧠 1,396+ open-source/commercial models: FLUX.2 Pro, GPT Image 2, Kling Video, Nano Banana 2, ElevenLabs TTS, MiniMax and more
- ✓⚡ <1 second cold start (fastest in market), per-second billing, auto-scale to zero when idle
- ✓🔌 OpenAI-compatible API + WebSocket Streaming + REST, unified interface across all models
- ✓🧩 Covers text-to-image, image-to-video, text-to-video, TTS, Audio, LLM, 3D and more modalities
- ✓💰 No monthly fees, no subscriptions, prepay $10+ and pay per model at exact usage rate
- ✓🏠 Python SDK (pip install fal) + TypeScript SDK + REST API, full language support
⚠️ Cons
- ×No free API tier (vs OpenRouter's free models, Replicate's trial credits)
- ×Model pricing fragmentation: 1,396+ models each with their own price, no unified token-based pricing table
- ×No built-in Guardrails / Prompt Management / Observability (vs Portkey or OpenRouter)
- ×FLUX inference pricing is typically 5-15% more expensive than BFL direct
- ×Real-time API docs lack detailed request/response schemas for some models
🎯 Best For
Developers and teams needing multimodal AI inference (image, video, audio, 3D); teams wanting access to 1,000+ models on one platform without managing multiple API keys; cold-start-sensitive inference scenarios (<1s fastest in market); rapid prototyping and per-second-billed lightweight production deployment
💰 Pricing & Plans
| Model | Category | Price | Notes |
|---|---|---|---|
| FLUX.2 [pro] | Image (T2I) | $0.03 1st MP + $0.015/extra MP | Black Forest Labs flagship; kontext/edit variants available |
| Nano Banana 2 | Image (T2I/I2I) | $0.08 / image | Google Gemini image model; 12 runs per $1 |
| Nano Banana Pro | Image (T2I/I2I) | $0.15 / image | Pro tier; double rate at 4K output |
| gpt-image-1.5 (GPT Image 2) | Image (T2I/edit) | Token-based: $5 in / $10 out per 1M (text) | OpenAI image gen; image tokens $8 in / $30 out per 1M |
| Kling Video v3 [pro] | Video (T2V/I2V) | $0.112–0.168 / sec | Kuaishou Kling 3.0; audio on adds ~50% |
| Veo 3.1 Fast | Video (T2V/I2V) | $0.10–0.15 / sec | Google video gen; fast tier for high throughput |
| Veo 3.1 Lite | Video (T2V/I2V) | $0.03–0.05 / sec | Budget tier; audio adds on |
| ElevenLabs TTS v3 | TTS | per-character (ElevenLabs rates) | Speech synthesis routed through ElevenLabs |
| MiniMax Music 2.6 | Music | $0.05 / sec | Text-to-music; $1 ≈ 20s of music |
🔧 API & Developer Experience
- •fal-serverless Engine: fal's proprietary inference runtime executes thousands of open and closed models with sub-second cold starts; auto-scales to zero when idle, so you pay only for actual compute, not for reserved capacity.
- •Async Queue + Webhooks: Generate endpoints run synchronously or via an async task queue with webhook callbacks and status polling — ideal for batch rendering jobs without holding open connections.
- •Single REST API: One base URL (api.fal.ai) covers image, video, audio, 3D, and LLM endpoints with consistent JSON schemas; each model exposes its own typed request/response contract in the docs.
- •Client SDKs: Official fal-client SDKs for Python and JavaScript/TypeScript, plus REST and WebSocket support; playground lets you test prompts before writing code.
- •Per-Second Billing: Open-source models are billed per second with sub-second granularity; image models per megapixel/image and video models per second — no monthly commitments or tier thresholds.
- •Fine-Tuning & Training Endpoints: FLUX LoRA / Kontext trainers, Wan image trainers, and Z-image trainers enable custom model fine-tuning priced per 1,000-step run.
- •OpenAI-Compatible Chat: For LLM tasks, fal exposes OpenAI-compatible chat completions so existing tool-calling and agent pipelines migrate with just a base-URL and key swap.
⚡ fal-serverless & Multi-Model Routing
fal.ai's signature value is fal-serverless, a proprietary inference runtime that bundles thousands of open and commercial models — FLUX, Nano Banana, Kling, Veo, Seedream, GPT Image, MiniMax, ElevenLabs — behind a single API. Its two defining traits are speed and cost control: cold starts under one second and auto-scaling to zero mean the platform is markedly faster than GPU-cloud peers like Replicate, while per-second/per-megapixel billing with no middleman markup keeps unit costs 30–50% below direct multi-provider integration. Beyond generation, fal exposes fine-tuning endpoints (FLUX LoRA/Kontext trainers), 3D asset pipelines, background removal, upscaling, and OpenAI-compatible chat completions for LLM work, making it a one-stop gateway for teams that A/B test many models in production rather than standardizing on a single vendor's lineup.
🌐 China Access & Latency
fal.ai is deployed primarily on AWS US-East and does not maintain mainland China edge nodes, so developers inside the GFW typically need a stable VPN or overseas VPS to reach api.fal.ai reliably; latency to the US East region from mainland cities is usually 250–500ms per request on top of model generation time, which is acceptable for turn-based generation but less ideal for real-time interactive workloads. Some Chinese teams instead access fal's catalog indirectly through domestic aggregators that re-sell the same endpoints with local billing and lower-latency routing. Because fal's per-second media generation is inherently asynchronous, moderate latency rarely blocks production use — but teams shipping to mainland users should plan for a proxy or a domestic relay in their architecture. For non-China regions, fal's latency is excellent and its model breadth makes it a strong default for multi-model media workloads.