Published Updated
In one or two sentences, before anything else.
fal is a San Francisco API platform, run by Features & Labels, Inc., that sells per-output access to more than 1,000 image, video, audio and 3D models on one prepaid balance. Its open models are cheap, but Google partner models such as Veo 3.1 Fast and Nano Banana Pro cost more on fal than on Google's own API.
What it can do, one sentence each, and whether we actually tested it.
Open-weight models that fal runs on its own GPUs (FLUX, Wan, Z-Image Turbo) and partner models such as Veo 3.1, Nano Banana Pro, Kling 3.0 and Seedance 2.0, all behind one key and one prepaid balance.
Model count from fal's homepage; partner models carry a Partner tag on their pages.
Submit, status, result and cancel endpoints with webhooks, plus a synchronous fal.run. fal says server errors are never billed and queued requests are never dropped; requests over your concurrency limit wait in the queue.
fal's docs; client errors such as 422 may be billed if GPU time was used.
mcp.fal.ai/mcp-relay lets Claude, Claude Code, Codex CLI, Cursor and ChatGPT search models, read prices and schemas, run models, manage queued jobs and upload files, signed in by OAuth.
Per fal's MCP docs; generations spend fal credits and fal says client approval prompts add no server-side approval gate.
17 skills for model routing, prompting and production workflows, installed by the genmedia CLI into .claude/skills or .agents/skills.
Published in the fal-ai-community GitHub organization; the repositories do not call themselves official.
Rent GPUs to deploy custom models next to the hosted ones: H100 at $4.50 an hour list ('as low as' $2.49), B200 at $7.99, B300 at $12.99.
fal's pricing page.
Accounts start at 2 concurrent requests; the limit rises with the last four weeks of paid invoices to 40 self-serve, and higher through sales.
fal's docs; some high-demand models add their own per-endpoint limits.
Official Python and JavaScript clients plus Swift, Kotlin/Java and Dart; a Vercel AI SDK provider for images, LiteLLM's fal_ai image route, and a Hugging Face Inference Provider listing.
The AI SDK and LiteLLM integrations are documented by Vercel and LiteLLM, not by fal.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
A 5-second, 720p MiniMax H3 video generated from the same scheduled prompts on fal, WaveSpeedAI and MiniMax's own API.
What happened
fal's median end-to-end time was 12.5 s (p95 36.5 s) at $3.6 per minute of video, against 123.2 s on MiniMax's API and 123.9 s on WaveSpeedAI at $6 per minute. The same benchmark has fal slow on Wan 2.2 A14B (271.6 s against Replicate's 85.2 s), so speed depends on the endpoint.
Case 2
Input
96 FLUX.1 Schnell and Dev runs and 40 SDXL runs, sent one at a time and in bursts of 4 and 8, timing queue, run and total time.
What happened
Replicate was faster on FLUX (about 1.40 s against fal's 3.49 s on Dev), mostly because of fal's queue time; fal was far faster on SDXL under load (about 3.3 s against 21.5 s for an 8-request burst). The site's operator is unnamed, the sample is small and some links are affiliate links.
Case 3
Input
Claude Code connected to https://mcp.fal.ai/mcp-relay, asked to find a video model, check its price and submit a job.
What happened
The server's tools cover search_models, get_pricing, submit_job, check_job and get_job_result; run_model waits only 45 seconds, so longer jobs go through the queue tools, and every resubmission is a new billable job. A Hacker News user said on October 14, 2026 that they make narrated videos with ChatGPT and the fal MCP server for about $3 to $5 each.
What it does badly, with the evidence, and what we could not settle.
Weak
Google's models cost more on fal than on Google's API.
Evidence: Nano Banana Pro is $0.15 per 1K or 2K image on fal against Google's $0.134, and $0.30 at 4K against $0.24; Veo 3.1 Fast with audio is $0.15 a second against Google's $0.10 at 720p (fal's model pages and Google's Gemini API pricing, October 21, 2026).
Most expensive of the four media services on Seedance 2.0.
Evidence: $0.3034 a second at 720p with audio on fal, against $0.24 on WaveSpeedAI, $0.205 on Kie.ai and $0.18 on Replicate (each service's model page, October 21, 2026).
Slow on some open models.
Evidence: Artificial Analysis' three-day medians (October 21, 2026): Qwen-Image 40.1 s on fal against 8.0 s on Replicate and 24.8 s on WaveSpeedAI; FLUX.2 [dev] 15.6 s on fal's standard endpoint against 6.4 s; Wan 2.2 A14B video 271.6 s against 85.2 s.
Low starting limits and credits that expire.
Evidence: New accounts start at 2 concurrent requests. Purchased credits expire 365 days after purchase, promotional credits after 90 days, and credits are non-refundable except at fal's discretion (fal's docs and terms).
Account and support complaints.
Evidence: Trustpilot: 2.4 out of 5 from 32 reviews, 69% one star, including an account locked after adding money with no reply for over a week (September 14, 2026) and unexpected charges (September 8, 2026); fal replied to 13% of negative reviews. A Hacker News user (May 12, 2026) reports minutes-long downloads from fal's CDN.
Not for buyers whose bill is mostly Veo or Nano Banana Pro, or for chat and coding-model apps.
Evidence: Google's own API is cheaper for its models, and fal's Model APIs serve media only. Teams after the lowest per-video price or an LLM endpoint should look elsewhere.
Uncertain
Minimum top-up and accepted payment methods.
Why unsure: fal's public docs and pricing page do not state them; they show only in the logged-in billing dashboard.
Size of any sign-up credit.
Why unsure: fal does not state an amount; third-party pages claim anywhere from $1 to about $20 and contradict each other.
Whether prompts and outputs are used for training.
Why unsure: The privacy policy (July 22, 2026) says prompts and uploads are used to provide, maintain and improve the service, and is silent on training outside enterprise contracts.
Real uptime.
Why unsure: The homepage claims 99.99%; the status page shows Model API at 99.97% over an unlabelled window and lists 25 incidents between September 2025 and September 2026.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | Features & Labels, Inc. (fal), San Francisco, California, US; founded 2021 | Founding year and founders per Contrary Research. Source: fal.aichecked |
| Price against the maker's list | At or above Google's list for partner models: Nano Banana Pro $0.15 per image vs $0.134; Veo 3.1 Fast with audio $0.15/s vs $0.10/s at 720p; Veo 3.1 at 1080p the same $0.40/s | Google's prices from its Gemini API pricing page, updated October 19, 2026. Source: fal.aichecked |
| Platform fee | None stated; any markup is inside each model's price | Source: fal.aichecked |
| Minimum top up and payment | Prepaid credits bought in the billing dashboard; minimum and card types not published | Source: fal.aichecked |
| Free tier | No standing free tier; credits granted case by case expire after 1 week to 1 year | Source: fal.aichecked |
| Models | 1,000+ image, video, audio and 3D models; no chat models on the Model APIs | Source: fal.aichecked |
| API formats | fal's own queue REST API (submit, status, result, cancel, webhooks) and synchronous fal.run; not OpenAI or Anthropic compatible | Source: fal.aichecked |
| MCP | Official hosted server at mcp.fal.ai/mcp-relay (OAuth): model search, prices, runs, queued jobs, uploads; docs-only server at fal.ai/docs/mcp | Source: fal.aichecked |
| Agent Skills | 17 skills in the fal-ai-community GitHub organization, installed by the genmedia CLI; not labelled official | Source: github.comchecked |
| Coding agents | MCP setup documented for Claude Code, Codex CLI, Cursor, Claude and ChatGPT | Source: fal.aichecked |
| Data retention | Outputs on fal's CDN for at least 7 days by default, configurable per request; SOC 2 claimed on the homepage | Source: fal.aichecked |
| Status page | status.fal.ai: Model API at 99.97% on October 21, 2026; homepage claims 99.99% | Source: status.fal.aichecked |
| Credit expiry and refunds | Credits expire 365 days after purchase (promotional after 90 days); non-refundable except at fal's discretion; server errors not billed | Source: fal.aichecked |
| Website visits, September 2026 (fal.ai) | 3.27 million (+21.9% on August); 12th of 20 on AICPB's AI cloud platforms list | Source: mp.weixin.qq.comchecked |
Answer first, then why, an example, the limits and what to do next.
fal is not cheaper than Google for Google's own models. Nano Banana Pro costs $0.15 per 1K or 2K image on fal against $0.134 on Google's API, and Veo 3.1 Fast with audio $0.15 a second against $0.10 at 720p; Veo 3.1 at 1080p is the same $0.40 a second on both.
Limits
Open-weight models are where fal is cheap: Artificial Analysis lists FLUX.1 [schnell] at $3 per 1,000 images on fal, Replicate and WaveSpeedAI alike.
fal is a real, venture-backed company with large named customers, but its terms are strict on credits and refunds. Contrary Research lists a Series D at a $4.5 billion valuation led by Sequoia, and fal names Canva and Adobe as customers. Credits expire after 365 days and are non-refundable, and Trustpilot scores it 2.4 out of 5 with complaints about locked accounts and slow support.
Why
The homepage claims SOC 2 compliance, and the privacy policy does not say whether prompts are used for training.
fal runs an official hosted MCP server at mcp.fal.ai/mcp-relay. Claude, Claude Code, Codex CLI, Cursor and ChatGPT sign in with OAuth and can search models, check prices, run models, manage queued jobs and upload files; a separate docs-only server searches fal's documentation without spending credits.
Limits
Generations spend your credits, run_model waits only 45 seconds, and fal says client approval prompts add no server-side approval gate.
fal has no standing free tier, only credits granted case by case. Its FAQ says such credits expire after one week to one year depending on the grant, and its terms say promotional credits expire after 90 days. fal does not state a sign-up amount.
Neither fal nor Replicate is faster across the board in Artificial Analysis' benchmarks. fal was quicker on FLUX.2 [klein] 9B (1.5 s against 2.1 s) and on MiniMax H3 video, which Replicate does not serve; Replicate was quicker on Qwen-Image (8.0 s against 40.1 s), FLUX.2 [dev] (6.4 s against 15.6 s) and Wan 2.2 A14B video (85.2 s against 271.6 s).
Limits
Three-day medians read on October 21, 2026, on open-weight models only.
The closest alternatives to fal are Replicate, WaveSpeedAI and Kie.ai, plus the makers' own APIs. Replicate charges the same as fal for Veo 3.1 and Nano Banana Pro but twice as much for Kling 3.0; WaveSpeedAI matches fal on Kling 3.0 and is slightly cheaper on Nano Banana Pro; Kie.ai is far cheaper but does not say how it gets access to the models.
Next step
For chat and coding models, an LLM router such as OpenRouter fits better.
This page is built from fal's own site, docs, terms and status page, Google's Gemini API price list, the model pages of Replicate, WaveSpeedAI and Kie.ai, Artificial Analysis' provider benchmarks, a fal-versus-Replicate test by AI Photo Labs, and user reports on Trustpilot and Hacker News, all checked on October 21, 2026. fal's claims are labelled as its own.
According to AICPB's traffic rankings, fal.ai had 3.27 million visits in September 2026, up 21.9% on August: 12th among AI cloud platforms.
fal sells one API key and one prepaid balance for more than 1,000 image, video, audio and 3D models, according to its homepage. Some are open-weight models that fal runs on its own GPUs, such as FLUX, Wan and Z-Image Turbo. Others are partner models with a "Partner" tag, such as Google's Veo 3.1 and Nano Banana Pro, Kling 3.0 and ByteDance's Seedance 2.0. On October 19, 2026 fal launched FLUX 3 Max video with Black Forest Labs. The Model APIs are media only: fal is not a router for chat models.
Prices are per output (per image, per megapixel, per second of video) and listed on each model page. fal also rents serverless GPUs for teams that deploy their own models, from $4.50 an hour list for an H100.
The company is Features & Labels, Inc. of San Francisco, founded in 2021. Contrary Research lists a $140 million Series D led by Sequoia at a $4.5 billion valuation (December 2025), and fal said in May 2026 that it serves more than 2.5 million developers, Canva and Adobe among them.
For open-weight models fal is cheap: Artificial Analysis lists FLUX.1 [schnell] at $3 per 1,000 images on fal, Replicate and WaveSpeedAI alike. For partner models fal charges at or above the maker's list price. Prices in US dollars, from each service's model or pricing page and Google's Gemini API pricing (updated October 19, 2026):
| Model and output | Google's list | fal | Replicate | WaveSpeedAI | Kie.ai |
|---|---|---|---|---|---|
| Veo 3.1 Fast, 8 s at 720p with audio | 0.80 | 1.20 | 1.20 | 1.20 | 0.30 |
| Veo 3.1, 8 s at 1080p with audio | 3.20 | 3.20 | 3.20 | 3.20 | 1.275 |
| Nano Banana Pro, one 1K or 2K image | 0.134 | 0.15 | 0.15 | 0.14 | 0.09 |
| Nano Banana Pro, one 4K image | 0.24 | 0.30 | 0.30 | 0.24 | 0.12 |
| Kling 3.0 Standard, per second without / with audio | not confirmed | 0.084 / 0.126 | 0.168 / 0.252 | 0.084 / 0.126 | 0.07 / 0.10 (720p) |
| Seedance 2.0, per second at 720p | not confirmed | 0.3034 (with audio) | 0.18 | 0.24 | 0.205 |
Veo 3.1 Fast costs 50% more on fal than on Google's API, and Nano Banana Pro about 12% more (25% at 4K); Veo 3.1 at 1080p matches Google. Kling's and BytePlus's own prices could not be confirmed, so the last two rows compare the four services only: fal ties WaveSpeedAI on Kling 3.0 and is the most expensive on Seedance 2.0 (Replicate's page does not say whether audio is included). Kie.ai charges per Veo clip, calls the audio experimental and does not say how it gets access.
fal lists no platform fee; any markup sits inside each model's price.
Artificial Analysis, an independent benchmarking firm, runs the same prompts on several providers every day and publishes the median end-to-end time, download included. Its three-day medians, read on October 21, 2026:
| Model | fal | Replicate | WaveSpeedAI |
|---|---|---|---|
| FLUX.1 [schnell] image | 3.2 s | 2.2 s | 11.0 s |
| FLUX.2 [klein] 9B image | 1.5 s | 2.1 s | 7.4 s |
| FLUX.2 [dev] image | 15.6 s | 6.4 s | not listed |
| Qwen-Image | 40.1 s | 8.0 s | 24.8 s |
| MiniMax H3, 5 s video at 720p | 12.5 s | not listed | 123.9 s |
| Wan 2.2 A14B, 5 s video at 720p | 271.6 s | 85.2 s | not listed |
fal was the fastest provider on MiniMax H3, about ten times quicker than MiniMax's own API (123.2 s) and cheaper per minute ($3.6 against WaveSpeedAI's $6). It was slow on Qwen-Image and on its standard Wan 2.2 and FLUX.2 [dev] endpoints (its "Regular Accel" FLUX.2 [dev] endpoint took 8.2 s). Only open-weight models are covered, so nothing here measures Veo or Nano Banana Pro on fal.
AI Photo Labs, a comparison site whose operator is not named, timed fal against Replicate on March 30, 2026. Replicate was faster on FLUX (about 1.40 s against 3.49 s on FLUX.1 Dev), mostly because of fal's queue time; fal was far faster on SDXL in a burst of eight requests (about 3.3 s against 21.5 s). The sample is small and the site carries some affiliate links.
fal's own claim that FLUX 3 Max renders a 10-second clip in 12.45 s against 277.80 s on Seedance 2.5 is a vendor benchmark, reported by VentureBeat on October 19, 2026.
Trustpilot shows 2.4 out of 5 from 32 reviews, 69% one star and 25% five star. Five-star reviewers in September 2026 praise the model range, batching and prices. One-star reviewers report an account locked after adding money with no reply for over a week (September 14, 2026), unexpected charges (September 8, 2026) and a LoRA that did nothing (October 19, 2026). fal had replied to 13% of negative reviews, usually after more than a month.
On Hacker News, a user whose team uses fal heavily in production said on May 12, 2026 that downloading a short video from fal's CDN sometimes takes minutes. Another, who runs the GenAI Showdown benchmark site, called fal "Highly recommended" on July 27, 2026.
fal runs an official hosted MCP server at https://mcp.fal.ai/mcp-relay. Per its docs, clients sign in with OAuth and get tools to search models, read schemas and prices, run a model, track queued jobs and upload files. Setup is documented for Claude, Claude Code, Codex CLI, Cursor and ChatGPT. A docs-only server at https://fal.ai/docs/mcp searches the documentation without spending credits. Generations bill your credits, and fal notes that client approval prompts "do not add a server-side approval gate."
For Agent Skills, the fal-ai-community GitHub organization publishes 17 skills (model routing, prompting, marketing and others), installed by the genmedia CLI into .claude/skills/ or .agents/skills/. The repositories do not call themselves official, though fal's MCP docs mention the genmedia CLI.
fal's API is its own queue format, not OpenAI or Anthropic compatible, so coding agents reach it through MCP, the CLI or the SDKs. Official clients exist for Python, JavaScript, Swift, Kotlin/Java and Dart. The Vercel AI SDK has a fal provider for images, LiteLLM has a fal_ai/ image route, fal is a Hugging Face Inference Provider, and it publishes an llms.txt index of its docs.
Replicate charges the same as fal for Veo 3.1 and Nano Banana Pro, twice as much for Kling 3.0 and less for Seedance 2.0, and was faster on several open models; it also lets you deploy your own model with Cog. WaveSpeedAI matches fal on Kling 3.0, is slightly cheaper on Nano Banana Pro, accepts Alipay and WeChat Pay, and was slower on most shared image models. Kie.ai is far cheaper on Google's models but has no status page and does not explain its access. For chat and coding models, an LLM router such as OpenRouter fits better.
Skip fal if your bill is mostly Veo or Nano Banana Pro: Google's own API is cheaper. Skip it for LLM-first apps, which it does not serve. And if you buy credits in bulk, plan to spend them within a year, because they expire and are not refunded.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
WaveSpeedAI is an API platform, run by a Singapore company and a Hong Kong company named in its terms, that sells 1,000+ image, video, audio and 3D models plus an LLM service on one balance. It matches fal on Kling 3.0, charges above Google's list for Veo 3.1 Fast and accepts Alipay and WeChat Pay.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Replicate is an API for running AI models, owned by Cloudflare since a $57.4 million acquisition closed on December 1, 2025. It sells 100+ official models such as Veo 3.1, Nano Banana Pro and Kling 3.0 per output, and 50,000+ community models by the second of GPU time.
Eden AI is a Lyon-based API gateway, run by Eden AI SAS, that puts about 500 language models and 250 'expert' models for OCR, speech, translation and image work behind one key and one prepaid balance. It passes provider prices through and, since January 2026, adds a 5.5% platform fee on self-serve credits.
CometAPI is a prepaid API service that sells text, image, video and audio models through one key at, by its own rule, 80% of the makers' list prices. Its terms and privacy pages name no company and use Hong Kong law, and it says it reaches the low prices through bulk credit buying and routing.