fal vs Replicate for Indie Developers (2026)
If you are adding image or video generation inside your product, fal and Replicate charge almost the same for the same hosted models: FLUX.1 [dev] is $0.025 and Nano Banana about $0.04 an image on both. The real differences are in the plumbing: fal starts new accounts at 2 concurrent requests and keeps outputs for at least 7 days, while Replicate allows 600 prediction requests a minute but deletes API outputs after an hour. Replicate is now owned by Cloudflare.
Key takeaways
- For the six models we checked on both sites, per-output prices match or differ by a fraction of a cent: FLUX.1 [schnell] $0.003, FLUX.1 [dev] $0.025, Seedream 4.5 $0.04, Kling 2.6 $0.07 per second without audio.
- fal bills hosted models per output or per megapixel and does not charge for server errors or queue time. Replicate bills official models per output and other public models per second of hardware time, and does not charge for failed public-model runs.
- fal new accounts start at 2 concurrent requests, rising to 40 as you buy credits. Replicate allows 600 prediction requests a minute, but only 6 a minute on granted credit with no card on file.
- Replicate deletes API prediction outputs after one hour by default. fal keeps generated media on its CDN for at least 7 days by default. Either way, copy files to your own storage.
- Cloudflare completed its acquisition of Replicate on 1 December 2025, per Cloudflare's SEC filing, and says existing Replicate APIs will keep working.
On this page 22
- TL;DR
- Who this is for
- fal vs Replicate at a glance
- How each one bills
- Same-model price comparison
- Worked example: 10,000 and 100,000 images a month
- GPU prices for custom deployments
- Custom models: Cog vs fal Serverless
- Limits, credits, queues and webhooks
- Licences, output rights and safety filters
- Who owns Replicate now
- Going direct to the model maker
- Decision table
- Setup checklist
- Frequently asked questions
- Is fal cheaper than Replicate?
- How long do fal and Replicate keep generated images?
- Can I use FLUX.1 [dev] images commercially?
- Did Cloudflare buy Replicate?
- Which is better for a custom or fine-tuned model?
- Ship it
- Sources
Verdict: For hosted image and video models, fal and Replicate cost almost exactly the same per output, so price should not decide it. Pick fal if you mainly call popular image and video models and want outputs that stay hosted for at least a week. Pick Replicate if you want a huge model catalogue, Cog-packaged custom models and high request limits from day one, and you are happy to copy every output to your own storage within the hour.
Every price, limit and policy below comes from fal's, Replicate's, Cloudflare's or Black Forest Labs' own pages, checked on 14 Nov 2026. The advice in between is ours and is marked as such.
TL;DR#
- Same model, same price. FLUX.1 [dev] is $0.025 on both, FLUX.1 [schnell] $0.003, Seedream 4.5 $0.04, Kling 2.6 $0.07 per second without audio (fal pricing, Replicate pricing).
- Billing units differ. fal charges hosted models per output or per megapixel. Replicate charges official models per output and most other public models per second of hardware.
- Limits differ more. fal starts you at 2 concurrent requests, up to 40 self-serve (fal concurrency). Replicate allows 600 prediction requests a minute (Replicate rate limits).
- Outputs expire. Replicate deletes API outputs after an hour (data retention). fal keeps them at least 7 days (fal FAQ).
- Ownership. Cloudflare acquired Replicate on 1 December 2025 (Cloudflare 10-Q).
Disclosure: ZeroRaised publishes this comparison and is a launch site for bootstrapped products. ZeroRaised has no commercial relationship with fal, Replicate, Cloudflare or Black Forest Labs and earns nothing if you sign up. Facts come from the linked sources as checked on 14 Nov 2026.
Who this is for#
This post is for a developer adding generation inside their own product, such as an avatar maker or a product photo tool, where you pay per API call. To rent raw GPUs instead, read our cheapest GPU cloud guide for indie AI developers.
fal vs Replicate at a glance#
| fal | Replicate | |
|---|---|---|
| Hosted model billing | Per output, per megapixel or per second of video, from prepaid credits | Official models per output; most other public models per second of hardware |
| Failed requests | Server errors (HTTP 500+) never charged; queue time free | Failed public-model runs not charged; private models bill instance time |
| Starting throughput | 2 concurrent requests, up to 40 self-serve | 600 prediction requests per minute |
| Output retention | At least 7 days on fal CDN by default | API outputs removed after 1 hour by default |
| Credit expiry | 365 days from purchase | 1 year from purchase, not refundable |
| Custom models | fal Serverless (fal.App, Docker, or an existing Cog image) | Cog containers, deployments |
| Official SDKs | JavaScript, Python, Swift, Kotlin/Java, Dart | Node.js, Python, Swift, Go |
| Async results | Queue API with fal_webhook | Webhooks with start, output, logs, completed events |
| Owner | Independent | Cloudflare (since 1 Dec 2025) |
Sources: fal pricing, fal FAQ, fal client libraries, Replicate pricing, Replicate billing, Replicate client libraries.
How each one bills#
fal prices Model APIs per output, drawn from prepaid credits: some per image, some per megapixel, video per second (fal pricing). Server errors are never charged, but client errors such as HTTP 422 "may still be charged if a runner spent GPU time processing the request before the error was detected" (fal FAQ).
Replicate prices official models (FLUX, Nano Banana, Kling) per output, while most other public models are "billed by run time" on the hardware's per-second rate (Replicate pricing). For public models, "if a run fails, we don't charge you", though cancelling an official model's run may still cost money (Replicate billing).
Our read: per-output pricing is what you want when you resell generations, because your cost per action is fixed.
Same-model price comparison#
These are the models where both vendors listed a price on 14 Nov 2026. We dropped any we could not verify on both.
| Model | fal | Replicate |
|---|---|---|
| FLUX.1 [schnell] | $0.003 per megapixel | $3 per 1,000 images ($0.003 each) |
| FLUX.1 [dev] | $0.025 per megapixel | $0.025 per image |
| FLUX.2 [pro], 1024×1024 output | $0.03 (first megapixel) | $0.015 per run + $0.015 per output megapixel |
| Nano Banana | $0.0398 per image | $0.039 per image |
| Seedream 4.5 (text to image) | $0.04 per image | $0.04 per image |
| Kling 2.6, per second of video | $0.07 audio off, $0.14 audio on | $0.07 without audio, $0.14 with audio |
Sources: fal pricing, fal FLUX.1 schnell, fal FLUX.1 dev, fal FLUX.2 pro, fal Nano Banana, fal Seedream 4.5, fal Kling 2.6 Pro, Replicate FLUX schnell, Replicate FLUX dev, Replicate FLUX.2 pro, Replicate Nano Banana, Replicate Seedream 4.5, Replicate Kling 2.6.
Notes on the table:
- Megapixels vs images. fal's FLUX.1 prices are per megapixel, Replicate's per image. fal's FLUX.2 page defines one megapixel as 1024×1024 and rounds up, so a 1920×1080 output costs $0.045. fal's FLUX.1 pages do not state the rounding rule, so test your exact size before you price a plan.
- Kling naming. fal lists the endpoint as Kling 2.6 Pro; Replicate lists
kwaivgi/kling-v2.6. fal's page gives the example that a 5-second clip with audio costs $0.70, and the Replicate rates give the same total. - Nano Banana Pro is $0.15 per standard image on both, with 4K at $0.30. fal charges $0.015 extra if web search is used.
Our read: the hosted-model market has converged on the model maker's list price. Choose on limits, retention and developer experience, not on fractions of a cent.
Worked example: 10,000 and 100,000 images a month#
This is illustrative arithmetic from the verified per-output prices above, for one model and no failed or retried calls.
| Monthly volume | Nano Banana on fal ($0.0398) | Nano Banana on Replicate ($0.039) | FLUX.1 [schnell], 1 MP, either ($0.003) |
|---|---|---|---|
| 10,000 images | $398 | $390 | $30 |
| 100,000 images | $3,980 | $3,900 | $300 |
The vendor gap is $8 at 10,000 images; the model gap is more than 13 times. If you charge per generation, our guide to setting usage limits for an AI product covers credits and caps so one heavy user cannot wipe out your margin.
GPU prices for custom deployments#
If you run your own fine-tune, LoRA pipeline or ComfyUI workflow, you pay for GPU time instead of per output.
| GPU | fal Serverless (list / as low as) | Replicate |
|---|---|---|
| T4 (16GB) | Not listed | $0.81/h |
| L40S (48GB) | Not listed | $3.51/h |
| A100 (80GB) | Not listed | $5.04/h |
| RTX PRO 6000 (96GB) | $4.00/h / $1.99/h | Not listed |
| H100 (80GB) | $4.50/h / $2.49/h | $5.49/h |
| H200 (141GB) | $6.00/h / $2.99/h | $5.49/h |
| B200 (192GB) | $7.99/h / $5.49/h | Not listed |
Sources: fal pricing, Replicate pricing.
Idle billing is the trap on both. Replicate private models bill "all time the instances are online", including setup and idle time, though fast-booting fine-tunes bill only active processing time (Replicate pricing). Failed runs on private models and deployments are billed for instance time as normal (Replicate billing). fal Serverless bills runners per second in the SETUP, IDLE, RUNNING, DRAINING and TERMINATING states, but not scheduling or image pulls (fal pricing).
Cold starts. Replicate's docs suggest a deployment with minimum instances set to 1 or more "to ensure that it is always running" (Replicate deployments). On fal, keep_alive keeps an idle runner warm after the last request, min_concurrency sets always-on runners, and setting it to 0 scales to zero so "an app with no traffic has no compute charge" (fal migration guide).
Our read: one always-on H100 for 720 hours is about $3,240 at fal's $4.50 list rate and about $3,950 at Replicate's $5.49. Before you have paying users, scale to zero or stay on per-output models.
Custom models: Cog vs fal Serverless#
Replicate packages models with Cog, "an open-source tool that lets you package machine learning models in a standard, production-ready container". fal Serverless takes a fal.App Python class, your own Dockerfile, or an existing Cog image with no code changes, though in that path outputs do not go to the fal CDN (fal migration guide). Our read: packaging with Cog keeps you free to move.
Limits, credits, queues and webhooks#
Concurrency and rate limits. fal: new accounts start at 2 concurrent requests, rising with paid invoices over the last four weeks to 40 self-serve. Extra requests wait in the queue and are never rejected for concurrency (fal concurrency). Replicate: 600 requests a minute to create predictions, 3,000 for other endpoints, but 6 a minute on granted credit with no payment method, and tighter limits as your balance runs low unless auto-reload keeps it above $20 (Replicate rate limits).
Our read: 2 concurrent video jobs will feel tight on launch day, so buy fal credit a few days ahead.
Prepaid credit. fal credits "expire 365 days from the date of purchase", and if your balance drops below your lock threshold, API requests are rejected (fal FAQ). Replicate credit is "valid for 1 year from the purchase date" and "not refundable", and auto-reload has a $5 minimum threshold and $15 minimum reload (Replicate prepaid credit).
Async jobs. Video takes longer than an HTTP request should wait. fal's queue API accepts a fal_webhook URL and sends a POST "when the request completes", and the queue retries requests that fail from temporary runner errors (fal queue submit). Replicate webhooks accept a webhook_events_filter of start, output, logs and completed, with output and logs throttled to at most once every 500ms (Replicate webhooks).
Output storage. This one bites people. Replicate: for API predictions, "all input parameters, output values, output files, and logs are automatically removed after an hour, by default" (Replicate data retention). fal: media is "available for at least 7 days by default", URLs are public by default, and you can set retention per request (fal FAQ). On either, your webhook handler should copy the file to your own bucket before you show it to the user.
Licences, output rights and safety filters#
The platform does not set your rights. The model licence does.
- FLUX.1 [schnell] is Apache 2.0 and "can be used for personal, scientific, and commercial purposes" (Replicate).
- FLUX.1 [dev] weights are under a non-commercial licence: using the model "for revenue-generating activity" is not a non-commercial purpose. But the licence also says "You may use Output for any purpose (including for commercial purposes)", except training a competing model (FLUX.1 dev licence). Replicate's page says "If you generate images on Replicate with FLUX.1 models and their fine-tunes, then you can use the images commercially", but not if you download the weights and run them yourself (Replicate FLUX dev). fal describes FLUX.1 [dev] as "suitable for personal and commercial use".
- Nano Banana outputs carry invisible SynthID watermarking, according to Replicate's model page.
- Other models: fal's FAQ says "Each model has its own license" and marks models as "Commercial use" or "Research only". Check each model page before you ship.
Safety filters are on by default. fal's FLUX.1 [dev] safety checker defaults to true; per the model's API reference, disabling it "requires account authorization", and flagged images come back black. Requests that hit a content filter return a non-retryable 422 content_policy_violation (fal FAQ). Replicate's FLUX.1 [dev] schema has a disable_safety_checker input that defaults to false.
Our read: show users a friendly "try a different prompt" message for filtered results instead of a black square, and do not automatically retry them.
Who owns Replicate now#
Cloudflare announced on 17 November 2025 that it "has agreed to acquire Replicate" (Cloudflare press release). Its 10-Q for the quarter to 30 September 2026 says that "On December 1, 2025, the Company acquired all of the outstanding shares of Replicate, Inc." for $57.4 million in cash (Cloudflare 10-Q). Cloudflare's blog told existing users: "Your APIs and workflows will continue to work without interruption" (Cloudflare blog).
Our read: a stability plus, but expect product changes, so keep generation code behind one adapter function.
Going direct to the model maker#
Black Forest Labs lists FLUX 1.1 [pro] at $0.04 an image, FLUX.1 [dev] at $0.025 and FLUX.2 [pro] at $0.03 for the first megapixel (BFL pricing). That matches the aggregators, so going direct buys early access to new releases, not savings. OpenAI prices GPT Image per token (gpt-image-1-mini is $8.00 per million image output tokens), so use its calculator for a per-image figure (OpenAI pricing).
Decision table#
This is our editorial advice, based on the facts above.
| Your situation | Our pick |
|---|---|
| You call popular image models (FLUX, Nano Banana, Seedream) | Either; prices match, so pick on developer experience |
| You need outputs to stay hosted for days | fal (at least 7 days by default) |
| You expect a launch spike on day one with a fresh account | Replicate (600 requests a minute) or pre-buy fal credit |
| You run long video jobs and want them queued, not rejected | fal (queued, never rejected for concurrency) |
| You want obscure community models | Replicate |
| You have a Cog model already | Replicate, or fal via its Cog path |
| You need cheap always-on H100 time for a fine-tune | fal Serverless ($4.50/h list) |
| You already build on Cloudflare Workers | Replicate |
| You only use FLUX and want new releases first | Black Forest Labs API |
Setup checklist#
- Test your exact output size on both and check the charge in each dashboard.
- Keep API keys on your server, behind a proxy route.
- Use queues and webhooks for anything slow, and copy outputs to your own storage there.
- Treat content-filter errors as user messages, not retries.
- Price generations above cost; see how to price an AI product.
- Put the provider behind one function so you can switch in an afternoon.
Frequently asked questions#
Is fal cheaper than Replicate?#
Not for the hosted models we compared. On 14 Nov 2026 FLUX.1 [schnell], FLUX.1 [dev], FLUX.2 [pro] at 1 megapixel, Seedream 4.5 and Kling 2.6 had the same listed price on both, and Nano Banana was $0.0398 on fal vs $0.039 on Replicate. fal is cheaper for raw H100 time on its serverless list ($4.50/h list, as low as $2.49/h, vs $5.49/h on Replicate).
How long do fal and Replicate keep generated images?#
Replicate removes inputs, outputs, output files and logs for API predictions after one hour by default. fal's FAQ says generated media is stored on its CDN for at least 7 days by default, and you can set retention per request with a header. In both cases, download outputs to your own storage straight away.
Can I use FLUX.1 [dev] images commercially?#
Black Forest Labs' FLUX.1 [dev] Non-Commercial License says you may use Output for any purpose, including commercial purposes, except training a competing model, but running the model itself for revenue needs a licence. Replicate's model page says images generated on Replicate can be used commercially, and fal labels FLUX.1 [dev] as suitable for personal and commercial use. FLUX.1 [schnell] is Apache 2.0.
Did Cloudflare buy Replicate?#
Yes. Cloudflare announced the deal on 17 November 2025, and its 10-Q for the quarter to 30 September 2026 says it acquired all outstanding shares of Replicate on 1 December 2025 for $57.4 million in cash. Cloudflare says existing Replicate APIs and workflows will continue to work.
Which is better for a custom or fine-tuned model?#
Both bill per second for custom compute. Replicate uses its open source Cog containers and bills private models for setup, idle and active time. fal Serverless bills runner time per second in setup, idle, running, draining and terminating states, can run an existing Cog image, and lists cheaper H100 time. Fast-booting fine-tunes on Replicate bill only active time.
Ship it#
Once your generation feature works, run through our guide on how to launch a bootstrapped app, then submit your product to ZeroRaised for free. For more no-VC build guides and comparisons, join the ZeroRaised newsletter.
Sources#
- fal pricing, last checked 14 Nov 2026
- fal Model APIs FAQ, last checked 14 Nov 2026
- fal concurrency limits, last checked 14 Nov 2026
- fal queue submit API, last checked 14 Nov 2026
- fal: migrate a Cog model to fal, last checked 14 Nov 2026
- fal client libraries, last checked 14 Nov 2026
- fal model pages: FLUX.1 schnell, FLUX.1 dev, FLUX.2 pro, Nano Banana, Nano Banana Pro, Seedream 4.5, Kling 2.6 Pro, last checked 14 Nov 2026
- Replicate pricing, last checked 14 Nov 2026
- Replicate billing and prepaid credit, last checked 14 Nov 2026
- Replicate rate limits, last checked 14 Nov 2026
- Replicate data retention, last checked 14 Nov 2026
- Replicate webhooks setup, last checked 14 Nov 2026
- Replicate deployments, last checked 14 Nov 2026
- Replicate client libraries, last checked 14 Nov 2026
- Replicate model pages: FLUX schnell, FLUX dev, FLUX.2 pro, Nano Banana, Nano Banana Pro, Seedream 4.5, Kling 2.6, last checked 14 Nov 2026
- Cog on GitHub, last checked 14 Nov 2026
- Cloudflare press release on Replicate, last checked 14 Nov 2026
- Cloudflare blog: Replicate is joining Cloudflare, last checked 14 Nov 2026
- Cloudflare Form 10-Q, quarter ended 30 Sep 2026, last checked 14 Nov 2026
- FLUX.1 dev Non-Commercial License, last checked 14 Nov 2026
- Black Forest Labs pricing, last checked 14 Nov 2026
- OpenAI API pricing, last checked 14 Nov 2026
Frequently asked questions
Is fal cheaper than Replicate?
Not for the hosted models we compared. On 14 Nov 2026 FLUX.1 [schnell], FLUX.1 [dev], FLUX.2 [pro] at 1 megapixel, Seedream 4.5 and Kling 2.6 had the same listed price on both, and Nano Banana was $0.0398 on fal vs $0.039 on Replicate. fal is cheaper for raw H100 time on its serverless list ($4.50/h list, as low as $2.49/h, vs $5.49/h on Replicate).
How long do fal and Replicate keep generated images?
Replicate removes inputs, outputs, output files and logs for API predictions after one hour by default. fal's FAQ says generated media is stored on its CDN for at least 7 days by default, and you can set retention per request with a header. In both cases, download outputs to your own storage straight away.
Can I use FLUX.1 [dev] images commercially?
Black Forest Labs' FLUX.1 [dev] Non-Commercial License says you may use Output for any purpose, including commercial purposes, except training a competing model, but running the model itself for revenue needs a licence. Replicate's model page says images generated on Replicate can be used commercially, and fal labels FLUX.1 [dev] as suitable for personal and commercial use. FLUX.1 [schnell] is Apache 2.0.
Did Cloudflare buy Replicate?
Yes. Cloudflare announced the deal on 17 November 2025, and its 10-Q for the quarter to 30 September 2026 says it acquired all outstanding shares of Replicate on 1 December 2025 for $57.4 million in cash. Cloudflare says existing Replicate APIs and workflows will continue to work.
Which is better for a custom or fine-tuned model?
Both bill per second for custom compute. Replicate uses its open source Cog containers and bills private models for setup, idle and active time. fal Serverless bills runner time per second in setup, idle, running, draining and terminating states, can run an existing Cog image, and lists cheaper H100 time. Fast-booting fine-tunes on Replicate bill only active time.