If you have generated an AI video through an app in 2026, there is a good chance it ran on fal.ai without you knowing. fal is a San Francisco company that hosts image, video and audio models behind one API, and on November 3, 2026 it announced a $1.3 billion Series F at a $16 billion valuation, double the $8 billion it was valued at in April. It also passed $800 million in annualized revenue this year. In this guide I explain what fal actually does, what its models and GPUs cost, how it compares with alternatives, and when running the same models on your own GPU makes more sense.
- What fal is: a generative media cloud. One API key gives you hundreds of image, video, audio and 3D models, plus serverless GPUs.
- The news: $1.3B round led by Bessemer with Nvidia, T. Rowe Price and Sound Ventures; $16B valuation; Series E (led by Sequoia) and F total $1.65B.
- Pricing: pay per output. Examples: FLUX.1 schnell $0.003 per megapixel, Nano Banana 2 $0.08 per image, MiniMax H3 Max video $0.03 per second, Kling v3 Pro $0.14 per second.
- GPUs: H100 from $2.49/h (list $4.50), H200 from $2.99/h, B200 from $5.49/h, B300 from $5.99/h.
- Who should use it: developers who need many models fast. Hobbyists with a 16 GB GPU can run many of the same open models free in ComfyUI.
What Is fal.ai?
fal was founded in 2021 by Burkay Gur (formerly a machine learning lead at Coinbase) and Gorkem Yurtseven (formerly at Amazon). It started as a fast way to run open image models like Stable Diffusion and FLUX, and grew into what it calls a generative media platform. There are two main ways to use it:
- Model APIs. Pick a model (FLUX, Nano Banana, Kling, MiniMax H3, Seedance, ElevenLabs voices and many more), send a prompt, get an image, video or audio file back. You pay per image, per second of video or per 1,000 characters.
- Serverless GPUs. Deploy your own model or ComfyUI workflow and pay for GPU time by the second, from RTX PRO 6000 cards up to B300.
Its selling points are speed (it is known for very fast inference on image models) and breadth: many closed models such as Kling and Google’s Nano Banana are available next to open ones, so apps can switch models without new contracts.
The $16 Billion Round: Key Facts
| Item | Detail |
|---|---|
| Round | Series F, $1.3 billion, announced November 3, 2026 |
| Lead investor | Bessemer Venture Partners |
| Other investors | Nvidia, T. Rowe Price, Sound Ventures, HOF Capital |
| Valuation | $16 billion, up from $8 billion in April 2026 |
| Combined with Series E | $1.65 billion (Series E led by Sequoia) |
| Revenue | More than $800 million annualized (per Bloomberg) |
| Use of money | Compute capacity and hiring |
CEO Burkay Gur said fal had to rate-limit customers over the summer because demand outran capacity. That is the backdrop to this raise: AI video exploded in 2026, and video eats far more GPU time than text. Competitors in inference hosting, including Baseten, Fireworks AI and Modal, are also raising large rounds, though TechCrunch notes that gross margins across these providers remain thin because compute is expensive.
fal.ai Model Pricing (Image, Video, Audio, 3D)
These are list prices from fal’s pricing page as of November 5, 2026. Prices scale with resolution and length, so check the model page before running large jobs.
Image models
| Model | Price |
|---|---|
| FLUX.1 [schnell] | $0.003 per megapixel |
| Z Image Turbo | $0.005 per megapixel |
| Nano Banana (generate and edit) | $0.0398 per image |
| Nano Banana 2 | $0.08 per image |
| Nano Banana Pro | $0.15 per image |
Video models
| Model | Price per second | 5-second clip |
|---|---|---|
| MiniMax H3 Max Turbo (text or image to video) | $0.015 | $0.075 |
| MiniMax H3 Max (text or image to video) | $0.03 | $0.15 |
| MiniMax H3 Max Reference to Video | $0.05 | $0.25 |
| Kling Video v2.5 Turbo Pro | $0.07 | $0.35 |
| Kling Video v3 Pro | $0.14 | $0.70 |
| Seedance 2 | $0.014 per 1,000 tokens | |
Audio and 3D
- ElevenLabs Eleven v3 and Multilingual v2 TTS: $0.10 per 1,000 characters; Turbo v2.5: $0.05.
- MiniMax Speech 2.8 HD: $0.10 per 1,000 characters; Turbo: $0.06. Gemini 3.1 Flash TTS: $0.05. xAI TTS v1: $0.015.
- ElevenLabs Sound Effects V2: $0.002 per second. ACE Step music: $0.0002 per second.
- 3D: Trellis $0.02 per generation, Trellis 2 $0.05, Hunyuan 3D Pro $0.015 per unit, Meshy 6 Preview $0.80 per generation.
For context on voices, my ElevenLabs alternatives guide compares quality and price, and my Kling AI guide covers Kling’s own plans.
fal.ai GPU Pricing
| GPU | VRAM | List price | As low as |
|---|---|---|---|
| RTX PRO 6000 | 96 GB | $4.00/h | $1.99/h |
| H100 | 80 GB | $4.50/h | $2.49/h |
| H200 | 141 GB | $6.00/h | $2.99/h |
| B200 | 192 GB | $7.99/h | $5.49/h |
| GB200 | 192 GB | $9.99/h | $5.89/h |
| B300 | 288 GB | $12.99/h | $5.99/h |
fal does not explain the conditions for the “as low as” rates on the page; they usually reflect committed or high-volume use. Serverless billing is per second, so you do not pay for idle time between requests. If you mainly want raw GPU hours, compare with my cheapest cloud GPU for AI list, and see H100 vs H200 vs B200 vs B300 for which card fits your model.
How to Start Using fal
- Create an account at fal.ai and add billing. The pricing page lists no free tier, so expect to buy credits. Third-party trackers note that paid credits expire after 365 days and promotional credits after 90 days.
- Create an API key in the dashboard and store it as
FAL_KEY. - Try a model in the browser playground first to check quality and the exact price.
- Call it from code with the official client:
pip install fal-client
import fal_client
result = fal_client.subscribe(
"fal-ai/flux/schnell",
arguments={"prompt": "a cozy reading nook, warm light, film photo"},
)
print(result["images"][0]["url"])Model IDs follow the fal-ai/<model> pattern; copy the exact ID from each model’s page, since names change between versions.
fal vs Replicate, Direct APIs and Local
| Option | Strength | Weakness |
|---|---|---|
| fal.ai | Fast inference, huge catalog incl. closed models, one bill | No free tier; prices can sit above going direct |
| Replicate, WaveSpeedAI and similar hosts | Similar catalogs, community models | Speed and price vary by model |
| Direct vendor APIs (Kling, Black Forest Labs, Google) | Sometimes cheaper, newest features first | One contract and SDK per vendor |
| Your own GPU (ComfyUI) | No per-image cost, full privacy, LoRAs and custom workflows | Upfront hardware, setup time, slower on big video models |
When Running Locally Is Cheaper
A quick rule of thumb. At $0.03 per second, a 5-second MiniMax H3 Max clip costs $0.15 on fal. A creator making 40 clips a day spends about $6 a day, or roughly $180 a month. MiniMax H3 open weights run in ComfyUI on a 16 GB card, so after the hardware cost the clips are nearly free apart from electricity, and you can use LoRAs that hosted APIs do not offer. On the other hand, if you need ten different models this week, or a closed model like Kling v3 or Nano Banana Pro, an API like fal is far simpler. My MiniMax H3 ComfyUI guide shows the local setup, and how much VRAM you need helps you size a GPU.
My Take
fal’s valuation doubling in seven months says more about AI video demand than about fal alone. Text models get cheaper every quarter, but video still needs a lot of GPU seconds per output, and someone has to own those GPUs. fal’s bet is that apps want one fast, reliable place to call every media model, and $800 million in run-rate revenue suggests they do. The risk is the one TechCrunch flags: thin margins, because fal pays for compute while vendors like Kling can undercut it on their own models. For developers, fal is an easy default for prototyping. For hobbyists and creators doing volume with open models, a local GPU still wins on cost.
Troubleshooting Common fal.ai API Errors
Most fal errors fall into a few groups: the key is missing or wrong, the input fails validation, the job takes too long, or your account has hit its concurrency or billing limits. fal returns a JSON error with a detail field, and request-level errors also carry an error_type (also sent as the X-Fal-Error-Type header), so read those before guessing.
| Problem | Likely cause | Fix |
|---|---|---|
| Unauthorized or forbidden response on every call | FAL_KEY not set in the process, a revoked key, or a raw HTTP call without the right header |
Export FAL_KEY in the same shell or server that runs your code. For raw HTTP, send Authorization: Key $FAL_KEY (the word “Key”, not “Bearer”). |
422 with value_error, missing or one_of |
A field is the wrong type, out of range or not sent | Read the loc and msg fields, then compare with the schema on the model page. Model IDs and parameters change between versions. |
422 with content_policy_violation |
The prompt or input image was flagged by fal’s or the partner model’s filter | Rephrase the prompt or change the image. This error is not retryable, so retrying the same input will fail again. |
422 with file_download_error |
fal could not fetch an input URL | Use a public URL with no login wall, or upload the file through the client’s storage helper first. |
504 request_timeout or generation_timeout |
Long video jobs or very large outputs | Use the queue (subscribe or submit) instead of a direct run call, shorten the clip or lower resolution, then retry. |
| 503 runner errors | No GPU runner was available at that moment | Retry with backoff. fal’s queue already retries 503 and 504 responses automatically, up to 10 attempts. |
429 concurrent_requests_limit |
Too many jobs running at once for your account | Retry with exponential backoff, or submit through the queue, where requests wait instead of failing. |
| Calls fail after working fine for days | Credits ran out or a card payment failed | Check your balance and invoices in the fal dashboard and top up. Before a big batch, estimate the cost from the model page so you do not run dry halfway. |
Why new accounts feel slow
fal’s documentation says new accounts start with a limit of 2 concurrent requests. Only jobs that are actually running count; queued jobs wait their turn and are never dropped. The limit rises automatically based on your paid invoices over the last four weeks, up to 40 on self-serve accounts, and higher limits go through sales. So if a batch of 50 images crawls on day one, that is expected. Buying credits is what lifts the cap.
A safe retry pattern
I’d only retry server-side failures (500, 502, 503, 504) and 429s, with an exponential backoff such as 2, 4, 8 and 16 seconds. Never auto-retry 4xx validation or content policy errors, because each attempt fails the same way. fal also sends an X-Fal-Needs-Retry header on model errors, which is the most reliable signal of whether a retry is worth it.
Frequently Asked Questions
Is fal.ai free?
No free tier is listed on fal’s pricing page. You pay per output or per GPU second. Occasional promotional credits exist but expire after about 90 days.
How much is fal.ai worth?
$16 billion after its $1.3 billion Series F in November 2026, double its April 2026 valuation of $8 billion.
Who owns fal.ai?
It is a private company founded by Burkay Gur and Gorkem Yurtseven, backed by investors including Bessemer, Sequoia, Nvidia and T. Rowe Price.
Is fal cheaper than Replicate?
It depends on the model. Compare the exact model page on both; for popular image models the gap is often small, while video pricing can differ more.
Why does my fal.ai request fail with a 422 error?
A 422 means fal rejected the input, not that the service is down. The response names the field and the reason, for example a value out of range, a missing required field, an image that is too small, or a prompt flagged by the content filter. Fix the field and resend; retrying unchanged will not help.
How many requests can I run at once on fal?
New accounts start at 2 concurrent running requests. The limit grows automatically as you pay invoices, up to 40 on self-serve plans, and extra requests wait in the queue rather than being dropped.
