What it is
fal (fal.ai) is a developer platform built for generative media. One API calls a large catalogue of image, video, audio and 3D models, including Google's Nano Banana family, FLUX, Kling, MiniMax and Seedance, and fal Serverless lets you run your own model on the same GPUs once the fal team approves your account.
Hosted models are billed per output: Nano Banana costs about $0.04 per image, Nano Banana 2 $0.08 and Nano Banana Pro $0.15, while Kling v2.5 Turbo Pro video is $0.07 a second. GPUs are listed from $4.50 an hour for an H100, falling to $2.49 at the lowest rate. fal raised $140 million led by Sequoia at a $4.5 billion valuation in December 2025, and its enterprise page names Adobe, Canva, Shopify and Quora among its customers.
Our review
fal earns its place by being early and legible. New media models tend to appear on it quickly, and the October 2026 pricing page lists models from Google, MiniMax, Kling, ByteDance, Black Forest Labs, OpenAI and ElevenLabs side by side, each with a per-image, per-second or per-character price. A product team can swap Nano Banana for Nano Banana 2 when quality shifts without signing a new contract or rewriting an integration.
The commercial momentum is real. Sacra estimates fal's annualised revenue at about $400 million in February 2026, up from roughly $25 million at the end of 2024, and the enterprise page says fal powers half of the image and video generation on Quora's Poe. That matters to a buyer because a supplier growing this fast is unlikely to vanish, though it also means terms and prices are still moving.
Two defaults deserve attention before launch. Media URLs fal returns are public to anyone holding the link until they expire, and request payloads are stored so they show in your dashboard unless you send the X-Fal-Store-IO: 0 header. Both can be changed, with access controls on outputs and per-request expiry, but neither is off by default. fal's promise never to train on customer data is worded for enterprise customers only; the public docs say nothing either way for everyone else.
Throughput is the other early surprise. New accounts can run only 2 requests at once, rising automatically to 40 as you buy credits, with anything beyond that negotiated with sales; extra requests queue rather than fail. Hosted prices are fair, but the top video models add up quickly at $0.14 a second, and an H100 at the $4.50 list rate is far cheaper than Fireworks' $8 for teams that only want raw GPUs. For language models fal is simply the wrong shop.
Where it earns its place
A product engineer at a 30-person photo-editing app
Add AI background replacement and edits without hosting models
Call Nano Banana through the queue API with webhooks; at about $0.04 an image, 200,000 edits a month is roughly $8,000. Set X-Fal-Store-IO: 0 and private ACLs, because uploads and outputs are otherwise kept and publicly linkable.
A growth lead at a DTC brand running paid social
Produce short product videos for ad testing each week
Generate 5-second Kling v2.5 Turbo Pro clips at $0.07 a second, about $0.35 each, so 500 variants cost around $175. Switching to Kling v3 Pro doubles that per clip.
An ML engineer at a startup with a fine-tuned image model
Serve a custom model without running a GPU fleet
Package it for fal Serverless with a Python app or Docker image once fal approves the account, then pay per GPU hour, $4.50 list for an H100 or less on discounted rates.
Before you commit
- Setup
- Sign up, buy credits and call a model with the Python or JavaScript client in minutes; deploying your own model on fal Serverless needs approval from the fal team.
- Integrations
- REST queue API with webhooks signed with ED25519, Python and JavaScript clients, a CDN for inputs and outputs, and a Platform API for deleting payloads.
- Data and privacy
- Payloads are stored by default and media URLs are public until expiry; per-request headers and ACLs change that. SOC 2 and SSO are listed for enterprise, and the no-training pledge covers enterprise customers; the report type is not published on the enterprise page.
- Cost at scale
- A workload of 100,000 Nano Banana Pro images a month costs about $15,000; a dedicated H100 running all month is about $3,300 at list price, less on discounted rates.
- Limits
- Concurrency starts at 2 requests and rises to 40 as you buy credits; above that needs sales. Free credits expire between one week and one year; bought credits last 365 days.
What's good
- New image, video and audio models land quickly
- Per-image, per-second and per-character prices on one page
- H100 from $2.49/hour on discounted rates
- Per-request controls for payload storage and media expiry
What's not
- Output URLs are public by default
- No-training promise published for enterprise customers only
- New accounts limited to 2 concurrent requests
- Not built for language model workloads
Best for
Product teams adding image or video generation to apps who want many models behind one API.
Alternatives
fal: common questions
- Is fal worth it?
- If image or video generation sits at the centre of your product, fal is where the new models arrive first and the per-output prices are easiest to budget. We rate it 4.4 out of 5.
- How much does fal cost?
- Per output, e.g. Nano Banana about $0.04/image, Nano Banana Pro $0.15/image, Kling v3 Pro $0.14/second of video, FLUX.1 [schnell] $0.003/megapixel. GPUs: H100 $4.50/hour list (from $2.49), B200 $7.99/hour (from $5.49). Purchased credits last 365 days. Enterprise by quote.
- What are the downsides of fal?
- Output URLs are public by default; no-training promise published for enterprise customers only; new accounts limited to 2 concurrent requests; not built for language model workloads.
- Who is fal best for?
- Product teams adding image or video generation to apps who want many models behind one API.
- How much does fal cost for a team or in production?
- A workload of 100,000 Nano Banana Pro images a month costs about $15,000; a dedicated H100 running all month is about $3,300 at list price, less on discounted rates.
- Is fal safe for company data?
- Payloads are stored by default and media URLs are public until expiry; per-request headers and ACLs change that. SOC 2 and SSO are listed for enterprise, and the no-training pledge covers enterprise customers; the report type is not published on the enterprise page.
- What are the best fal alternatives?
- The closest alternatives we have reviewed are Replicate, Modal, Together AI and Hugging Face.
Answers checked Oct 22, 2026.
Make fal? Show this rating on your site
