How does fal.ai pricing work?

How per-output model pricing works

The standard fal Model APIs charge you based on the work the model delivers rather than raw server uptime. You make an API call, receive a media payload, and fal debits your credit balance based on images, seconds, or tokens.

Image model pricing

Image models are billed either by image count or by output size measured in megapixels (MP). fal normalizes its catalog pricing to a baseline 1MP image, and any request rendered above 1MP is billed proportionally higher. Most listed models follow this per-output structure, though some other image models on the platform use GPU-based pricing depending on how their underlying architecture runs.

At the low end, FLUX.1 [schnell] (fal-ai/flux/schnell) costs $0.003 per megapixel, which gives you 333.33 megapixels per $1. Z Image Turbo (fal-ai/z-image/turbo) costs $0.005 per megapixel, yielding 200 megapixels per dollar.

Models priced per individual output follow different rates:

• Nano Banana (fal-ai/nano-banana and fal-ai/nano-banana/edit): $0.0398 per image (25.13 images per $1).
• Nano Banana 2 (fal-ai/nano-banana-2 and fal-ai/nano-banana-2/edit): $0.08 per image (12.5 images per $1).
• Nano Banana Pro (fal-ai/nano-banana-pro and fal-ai/nano-banana-pro/edit): $0.15 per image (6.67 images per $1).
• GPT Image 2 API (openai/gpt-image-2 and openai/gpt-image-2/edit): $1 per unit (1 unit per $1).

Video model pricing

Video models are billed by output unit, either per second or per whole video depending on the specific model. fal bases its normalized pricing comparisons on an estimated average generation of 5 seconds at 720p, though your actual bill will vary based on resolution, duration, prompt complexity, and model family.

MiniMax endpoints sit across three distinct price tiers:

• H3 Max Turbo Image to Video (minimax/h3-max-turbo/image-to-video) and Text to Video (minimax/h3-max-turbo/text-to-video) cost $0.015 per second (66.67 seconds per $1).
• H3 Max Image to Video (minimax/h3-max/image-to-video) and MiniMax H3 Max Text to Video (minimax/h3-max/text-to-video) run at $0.03 per second (33.33 seconds per $1).
• H3 Max Reference to Video (minimax/h3-max/reference-to-video) costs $0.05 per second (20 seconds per $1).

For other video engines, Kling Video (fal-ai/kling-video/v2.5-turbo/pro/image-to-video) costs $0.07 per second, giving you 14.29 seconds per dollar. The higher-tier Kling Video v3 Image to Video [Pro] (fal-ai/kling-video/v3/pro/image-to-video) and Text to Video [Pro] (fal-ai/kling-video/v3/pro/text-to-video) are priced at $0.14 per second, or 7.14 seconds per $1.

Bytedance models use a different metric entirely. Seedance 2 Reference to Video (bytedance/seedance-2.0/reference-to-video) and Seedance 2 Image to Video (bytedance/seedance-2.0/image-to-video) bill at $0.014 per 1,000 tokens, which works out to 71.43 units of 1,000 tokens per dollar.

Audio and 3D generation

Audio endpoints are billed per 1,000 characters or per generated second. ACE Step Prompt To Audio (fal-ai/ace-step/prompt-to-audio) costs $0.0002 per second, giving 5,000 seconds per $1. Elevenlabs Sound Effects V2 (fal-ai/elevenlabs/sound-effects/v2) costs $0.002 per second.

Text-to-speech models charge per thousand characters:

• xAI Text to Speech (xai/tts/v1): $0.015 per 1,000 characters (66.67 units per $1).
• ElevenLabs TTS Turbo v2.5 (fal-ai/elevenlabs/tts/turbo-v2.5) and Gemini 3.1 Flash Tts (fal-ai/gemini-3.1-flash-tts): $0.05 per 1,000 characters (20 units per $1). You can compare this with standalone API rates in our breakdown on How does ElevenLabs pricing work for the API?.
• MiniMax Speech 2.8 [Turbo] (fal-ai/minimax/speech-2.8-turbo): $0.06 per 1,000 characters.
• MiniMax Speech-02 HD (fal-ai/minimax/speech-02-hd), MiniMax Speech 2.8 [HD] (fal-ai/minimax/speech-2.8-hd), ElevenLabs TTS Multilingual v2 (fal-ai/elevenlabs/tts/multilingual-v2), and Eleven V3 (fal-ai/elevenlabs/tts/eleven-v3): $0.1 per 1,000 characters.

3D generation is billed in units, generations, or credits. Sam 3 (fal-ai/sam-3/3d-objects) charges $0.02 per unit, and Trellis (fal-ai/trellis) is $0.02 per generation, both delivering 50 units per dollar. Hunyuan 3D Pro Image to 3D and Text to 3D (fal-ai/hunyuan-3d/v3.1/pro/image-to-3d and text-to-3d) cost $0.015 per unit, Trellis 2 (fal-ai/trellis-2) costs $0.05 per unit, and Hunyuan Motion [1B] (fal-ai/hunyuan-motion) runs $0.08 per generation. Tripo H3.1 endpoints (tripo3d/h3.1/image-to-3d, text-to-3d, multiview-to-3d) charge $0.01 per credit, while Meshy 6 Preview (fal-ai/meshy/v6-preview/image-to-3d) costs $0.8 per generation.

fal Serverless and Compute GPU pricing

If you run custom inference pipelines or deploy proprietary checkpoints via fal Serverless and Compute, you pay for the underlying GPU capacity by the hour.

fal lists both a standard list price and discounted pricing.

The B300 offers the highest memory footprint at 288 GB of VRAM, with list pricing cut down to $5.99/h at volume. Enterprise customers can contact sales directly to negotiate volume discounts and set up invoice-based billing.

Credits, balance locks, and cold starts

Billing on fal operates primarily on a prepaid credit balance. You buy credits, and individual API calls draw down that balance.

Purchased credits expire 365 days from the date of purchase. That gives you a full year to burn through any prepayments you deposit on the platform. If you receive promotional grants, free credits, or coupons, those have variable expirations that range between 1 week and 1 year depending on the terms of the specific grant. You can review your current credit balance and exact expiration dates at any time inside the billing dashboard.

Be careful with your balance floor. When your credit balance drops below your account's lock threshold, your account is immediately locked and incoming API requests will be rejected. You must add fresh credits inside the billing dashboard to unlock it and resume processing traffic. Enterprise accounts on invoice-based billing are exempt from this automated lock behavior.

Cold starts are free. For standard Model API endpoints, you are billed strictly for active inference time. The platform does not charge you for container pull times or model loading delays, which prevents unpredictable billing spikes when cold containers spin up.

Server errors are free as well. Any request that fails with an HTTP 500+ status code due to a runner infrastructure issue is never charged. Client errors are different. If you submit bad parameters and hit an HTTP 422 error, fal may still charge your balance if the runner spent GPU processing time before the error was detected.

Concurrency and media retention rules

Every new account starts with a concurrency limit of 2 concurrent requests. That limit is tight. As you purchase credits, fal increases this ceiling automatically up to 40 concurrent requests.

Any requests submitted beyond your active limit are queued and processed sequentially as runner slots free up. If your application architecture requires higher throughput than 40 parallel executions, you have to contact fal sales to request a limit increase.

Generated images, videos, and audio files are hosted directly on fal's CDN. By default, they remain accessible for at least 7 days. If your workload handles sensitive data or you want to control storage lifecycles directly, you can override this retention window per request by passing the X-Fal-Object-Lifecycle-Preference header.

Discount structures have one major catch: Partner API models. Standard percentage discounts do not apply to Partner API models when called directly through the API, leaving those calls on standard pay-as-you-go pricing. However, subscription discounts from fal Agent Pro and Max do apply to Partner API models if you access them through the fal UI, Playground, Sandbox, or CLI.

Monitoring fal.ai spend with Costbrake

Because fal locks your API access the moment your prepaid balance dips below its lock threshold, an untracked surge in generation volume can silently knock your production services offline.

We built Costbrake to watch your cloud, AI, and SaaS platforms in one place. It tracks your billing pace across 23 providers - including fal.ai, OpenAI, Anthropic, AWS, RunPod, and Vast.ai - checking spending hourly on the Free plan and every 5 minutes on Pro and Team (ramping up to every 2 minutes for fal.ai during active spending spikes). You can read more about coordinating several inference services in our guide on how to monitor spend across OpenAI, Anthropic, Gemini and other AI APIs in one place.

Costbrake connects to fal.ai using a restricted read-only API key that can only query usage, keys, and team members. It cannot run inference. You can generate this credential by passing the billing:usage:read, auth:keys:read, and account:users:read permissions:

D=$(mktemp -d /tmp/costbrake.XXXXXX)
curl -s -X POST https://api.fal.ai/v1/keys -H "Authorization: Key $FAL_KEY" -H "content-type: application/json" \
  -d '{"alias":"costbrake","permissions":["billing:usage:read","auth:keys:read","account:users:read"]}' > "$D/key.json"

Send the resulting key string as apiKey without storing the secret, or build the key by hand at https://fal.ai/dashboard/keys with those three scopes. Because fal keys can only be revoked permanently, Costbrake does not provide a reversible killswitch (brake) for fal.ai. It focuses purely on usage tracking, spike detection, and budget alerts before your balance hits zero. You can connect it via agent setup or review plan details on Costbrake pricing.