In short
- On 5 October 2026 no host was cheapest for all six models: WaveSpeed won on Nano Banana Pro, Replicate on Nano Banana 2 and Seedance 2.0, fal on MiniMax H3, and WaveSpeed and fal tied on Kling 3.0.
- Replicate charged twice as much as WaveSpeed and fal for Kling 3.0 image-to-video, but less than both for Seedance 2.0 at 720p.
- The same model is billed per image, per second, per megapixel or per 1,000 tokens depending on the host, so always convert to the cost of one finished output.
- WaveSpeed credits never expire, while purchased fal credits and Replicate credit run out after a year.
- Replicate deletes API outputs after one hour by default, while WaveSpeed and fal keep files for about seven days.
On this page
- WaveSpeed vs fal vs Replicate: six models priced on one day
- Where each price comes from
- Why the units make AI hosts hard to compare
- Promotions change the price you see
- How each host bills you: credits, expiry and failed runs
- Rate limits and file storage for batch work
- LoRA training on WaveSpeed, fal and Replicate
- Which host to pick for an AI persona
- Questions people ask
WaveSpeed vs fal vs Replicate has no single winner. On 5 October 2026 we priced six models at the same settings on all three hosts. WaveSpeed was cheapest for Nano Banana Pro, Replicate for Nano Banana 2 and Seedance 2.0, and fal for MiniMax H3. WaveSpeed and fal tied on Kling 3.0. Check the model you use, because these prices move.
All three hosts sell the same popular models through an API and bill you per output. The price for one model can differ by a factor of two between them. Below: the price table, the unit traps, and the rules for credits, limits and files.
WaveSpeed vs fal vs Replicate: six models priced on one day
Each row is one model at one setting, read from each host’s own model page on 5 October 2026. The prices are list prices without promotions.
| Model and setting | WaveSpeed | fal | Replicate |
|---|---|---|---|
| Nano Banana 2, one image at 1K | $0.07 | $0.08 | $0.067 |
| Nano Banana Pro, one image at 1K | $0.14 | $0.15 | $0.15 |
| Kling 3.0 Pro image-to-video, 5 s, no sound | $0.56 | $0.56 | $1.12 |
| Kling 3.0 Std image-to-video, 5 s, no sound | $0.42 | $0.42 | $0.84 |
| Seedance 2.0 image-to-video, 720p, 5 s | $1.20 | about $1.51 | $0.90 |
| MiniMax H3 image-to-video, 768p, 5 s | $0.50 | $0.30 | $0.40 |
Bars are scaled within each group.
Nano Banana 2one image, 1K
Nano Banana Proone image, 1K
Kling 3.0 Pro5 s, no sound
Kling 3.0 Std5 s, no sound
Seedance 2.0720p, 5 s
MiniMax H3768p, 5 s
SourceWaveSpeed, fal and Replicate model pages, quoted belowchecked
Bold marks the cheapest host in each row, or the two that tie. Replicate says its Kling standard mode is 720p and pro is 1080p. The WaveSpeed and fal pages quoted below use the same Std and Pro names but give no resolution.
Two rows show the biggest gaps. Kling 3.0 costs twice as much on Replicate as on the other two hosts. Seedance 2.0 at 720p costs about 40% less on Replicate than on fal. For a persona that makes most of its video with one model, that difference adds up every month.
Here is a worked example: one month with 200 Nano Banana 2 images and 30 five-second Kling 3.0 Std clips, at the list prices above.
| Host | 200 images | 30 clips | Month |
|---|---|---|---|
| WaveSpeed | $14.00 | $12.60 | $26.60 |
| fal | $16.00 | $12.60 | $28.60 |
| Replicate | $13.40 | $25.20 | $38.60 |
Replicate is the cheapest host for the images and still the most expensive for this month. Swap Kling for Seedance 2.0 and the order changes. So work out your own mix before you pick a host. What an AI persona costs to start puts these numbers into a full budget.
Where each price comes from
Every price in the table is quoted below, with the date it was read. Prices on all three hosts change without notice, so the dated list on what generation costs is the place to check before you commit.
WaveSpeed prices
Pricing by resolution: 0.5k $0.045, 1k $0.07, 2k $0.105, 4k $0.14.
WaveSpeedAI — Nano Banana 2 Edit (opens in a new tab)checked
Cost per image: 1k $0.14, 2k $0.14, 4k $0.24. The page header showed “$0.126 per run” with a “10 % OFF” badge.
WaveSpeedAI — Nano Banana Pro Edit (opens in a new tab)checked
“Base rate: $0.56 per 5 seconds ($0.112 per second)”. “Sound surcharge: ×1.5 when sound is enabled”. Duration range 3 to 15 seconds.
WaveSpeedAI — Kling v3.0 Pro Image-to-Video (opens in a new tab)checked
“Base rate: $0.42 per 5 seconds ($0.084 per second)”. “Sound surcharge: ×1.5 when sound is enabled”.
WaveSpeedAI — Kling v3.0 Std Image-to-Video (opens in a new tab)checked
480p 5 s $0.60, 720p 5 s $1.20, 1080p 5 s $3.00. “Prices scale linearly with duration (4-15 seconds).” “Native audio generation is included”. The page header showed “$0.54 per run” with a “10 % OFF” badge.
WaveSpeedAI — Seedance 2.0 Image-to-Video (opens in a new tab)checked
Pricing is based on output resolution and duration: 768p $0.10 per second ($0.50 for 5 s), 2K $0.14 per second ($0.70 for 5 s).
WaveSpeedAI — MiniMax H3 Image-to-Video (opens in a new tab)checked
fal prices
“Your request will cost $0.08 per image.” “2K and 4K outputs will be charged at 1.5 times and 2 times the standard rate, respectively. 0.5K (512px) resolution outputs will be charged at 0.75 times the standard rate.” “Note: Pricing may change in the future.”
“Your request will cost $0.15 per image.” “4K outputs will be charged at double the standard rate.”
“For every second of video you generated, you will be charged $0.112 (audio off) or $0.168 (audio on)”. 5 s example: $0.56 with audio off, $0.84 with audio on.
fal — Kling Video v3 Pro Image-to-Video (opens in a new tab)checked
“For every second of video you generated, you will be charged $0.084 (audio off) or $0.126 (audio on)”.
fal — Kling Video v3 Standard Image-to-Video (opens in a new tab)checked
“Your request will cost $0.014 per 1000 tokens for 480p, 720p and 1080p.” “The number of tokens is given by (height of output video * width of output video * duration * 24) / 1024.” The pricing table lists the standard tier at $0.3024 per second; the text above it says $0.3034 per second for 720p. “Audio generation is included at no extra cost regardless of the generate_audio setting.”
fal — Seedance 2.0 Image-to-Video (opens in a new tab)checked
“Video costs $0.05 per second at 480p, $0.06 per second at 768p, $0.13 per second at 2K and $0.16 per second at 4K.”
For Seedance, the token formula gives 108,000 tokens for 5 seconds at 1280 × 720, which is $1.512. That matches the $0.3024 in fal’s table, not the $0.3034 in the text above it.
Replicate prices
Per output image: 1K $0.067, 2K $0.101, 4K $0.151.
Replicate — google/nano-banana-2 (opens in a new tab)checked
Per output image: 1K $0.15, 2K $0.15, 4K $0.30.
Replicate — google/nano-banana-pro (opens in a new tab)checked
Per second of output video: standard $0.168 without audio and $0.252 with audio; pro $0.224 without audio and $0.336 with audio. “mode: standard (720p) or pro (1080p)”.
Replicate — kwaivgi/kling-v3-video (opens in a new tab)checked
Per second of output video without a video input: 480p $0.08, 720p $0.18, 1080p $0.45. With a video input: 480p $0.10, 720p $0.22, 1080p $0.55.
Replicate — bytedance/seedance-2.0 (opens in a new tab)checked
“MiniMax H3 is billed per second of generated video. 768P outputs cost $0.08 per second, and 2K outputs cost $0.13 per second.”
Why the units make AI hosts hard to compare
The same model is sold per image on one host and per megapixel or per token on another. Convert every price to the cost of one finished output at your settings before you compare.
Z-Image-Turbo shows the trap well. It is a cheap image model on all three hosts:
“$0.005 per generated image”.
“Your request will cost $0.005 per megapixel.” Optional prompt expansion “adds $0.0025 per request”.
“Pricing is based on image resolution (height × width): Up to 0.5 MP (500,000 pixels): $0.0025. Up to 1 MP (1,000,000 pixels): $0.005. Up to 2 MP (2,000,000 pixels): $0.01”.
Replicate — prunaai/z-image-turbo (opens in a new tab)checked
A 1024 × 1024 image has 1,048,576 pixels. On WaveSpeed it costs $0.005 as one image. On fal it is 1.05 megapixels, so about $0.0052. On Replicate it is just over the 1 MP line, so by that wording it falls into the $0.01 band. The same square image costs twice as much on one host because of 48,576 pixels. At 1000 × 1000 it would cost $0.005 everywhere.
The units in the main table:
| Unit | Who uses it | What to watch |
|---|---|---|
| Per image | WaveSpeed, fal and Replicate for Nano Banana | Higher resolutions cost more, by different factors per host |
| Per second | All three for Kling, MiniMax H3 and Replicate’s Seedance | Sound can add 50% on Kling |
| Per 5 seconds | WaveSpeed’s Kling and Seedance tables | Same as per second, written another way |
| Per 1,000 tokens | fal’s Seedance 2.0 | Tokens grow with width, height and length |
| Per megapixel | fal’s Z-Image Turbo | A non-square or large image costs more than “one image” |
What the sources do not say
Pricing the same six models on all three hosts at the same settings
We wanted GPT Image 2.5 in the table, because it is the newest image model our readers ask about. We could not price it on fal.
GPT Image 2.5 is not in fal’s pricing table, and we found no price on its fal model page. fal’s own index says each model’s llms.txt holds its current price, but fal.ai/models/openai/gpt-image-2.5/llms.txt returned “not found” on 5 October 2026. So GPT Image 2.5 left the table.
Two more rows needed care. MiniMax H3 is sold as more than one product: WaveSpeed has the official H3 and a separate open-weights edition with its own prices, and fal has H3 and “H3 Max”. The table uses the official H3 on all three hosts. And fal’s Seedance 2.0 page gives two different per-second prices for 720p, $0.3024 and $0.3034, so we used its token formula.
You can check this yourself. Open fal.ai/pricing and search for “gpt-image-2.5”. Then open the MiniMax H3 pages on fal, WaveSpeed and Replicate and compare the resolutions each one offers.
Promotions change the price you see
All three hosts ran a promotion on 5 October 2026. A promotion can make the wrong host look cheapest for a week.
“Video costs $0.03 per second at 480p, $0.048 per second at 768p, and $0.096 per second at 1080p.” “Note: these are promotional rates, 40% off for a limited time. The discount ends October 15, after which 480p is $0.05/second, 768p is $0.08/second, and 1080p is $0.16/second.”
fal — MiniMax H3 Max Image-to-Video (opens in a new tab)checked
The page banner read “Wan 3.0 is 30% off this week”. “Most models are billed by the time they take to run.” “Some models are billed by input and output.” “You’ll find estimates for how much any model will cost you on the model’s page.”
WaveSpeed showed discount badges on several model pages, as quoted above. The table uses the list prices for that reason. Promotions end, and a batch you plan this week may run after the discount.
How each host bills you: credits, expiry and failed runs
All three hosts take prepaid credit. They differ on how long it lasts and on what happens when a job fails.
“Purchased credits expire 365 days from the date of purchase.” “Server errors (HTTP 500+) are never charged.” “Client-side errors like invalid inputs (HTTP 422) may still be charged if a runner spent GPU time processing the request before the error was detected.” New accounts “start at 2 concurrent requests. As you purchase credits, the limit increases automatically up to 40.” Generated media files are “available for at least 7 days by default”. Media URLs returned by fal “are publicly accessible — anyone with the URL can access the file until it expires.”
“Purchased credit is valid for 1 year from the purchase date and is not refundable.” “As soon as your balance hits zero, we will prevent any new work from starting”.
“For all models, if a run fails, we don’t charge you. If you cancel a run for an official model, you may still be charged.”
Failed requests, system errors and system timeouts are eligible for automatic credit reversal. A Sync Mode request that times out on the client side while the prediction completes “is not automatically refundable”.
WaveSpeedAI docs — Refund Policy (opens in a new tab)checked
Credits are purchased via top-up, “never expire”, and “are non-refundable but remain in your account”.
WaveSpeedAI docs — How Pricing Works (opens in a new tab)checked
Bronze, the default for new users: 5 predictions per minute and a max concurrency of 2. “Any successful single top-up below $1,000” moves an account to Silver: 500 per minute and 300 at the same time.
WaveSpeedAI docs — Account Levels & Rate Limits (opens in a new tab)checked
| WaveSpeed | fal | Replicate | |
|---|---|---|---|
| Credit expires | Never | 365 days after purchase | 1 year after purchase |
| Failed job | Refunded automatically | Server errors free; bad inputs may be charged | Not charged; a cancelled official-model run may be |
| Jobs at once for a new account | 2, then 300 after any top-up | 2, up to 40 as you buy credit | Limited by requests per minute instead |
| Generated files kept | Generally up to 7 days | At least 7 days by default | 1 hour for API runs |
WaveSpeed is the only one of the three whose credit never runs out. That matters if you top up once and generate in bursts. What WaveSpeed is good and bad at beyond price is in the WaveSpeedAI review. How to use WaveSpeed AI shows how to check a price and your balance before a batch.
Rate limits and file storage for batch work
For batches, two rules matter most: how many jobs run at once, and how long the output stays online.
“You can create predictions at 600 requests per minute.” “As you approach running out of credit, we apply stronger rate limits.” “If you have been granted credit and don’t have a payment method on file, you’ll also be rate limited to 1 request per second with a maximum of 6 requests per minute.”
“For predictions created through the API, all input parameters, output values, output files, and logs are automatically removed after an hour, by default.” “Data for predictions created through the web interface is kept indefinitely.”
Generated media files: retention “Up to 7 days”. “Generated media files are temporary and generally expire within 7 days.”
WaveSpeedAI docs — Data Retention Policy (opens in a new tab)checked
On Replicate, a script that saves its files the next morning finds nothing. Save every output in the same run that created it. On fal, output links are public by default, so treat them like a shared link until they expire.
LoRA training on WaveSpeed, fal and Replicate
All three hosts can train a character LoRALoRA A LoRA is a small file trained on images of one subject. It teaches a general image model to reproduce that specific face or style. Open the glossary for you, but the choice of trainers and the billing differ.
The collection lists 13 models, including the Z-Image LoRA Trainer, Flux-dev LoRA Trainer, Wan 2.2 Image LoRA Trainer, Wan 2.2 I2V LoRA Trainer and Qwen-Image LoRA Trainer.
WaveSpeedAI — Training Tools collection (opens in a new tab)checked
“Your request will cost $2 per training run (scales linearly with steps).”
“You’ll be billed per second for the time the training process takes to run. Trainings for the FLUX model run on x8 Nvidia H100 GPU hardware, which costs $0.012200 per second at the time of this writing.”
Replicate docs — Fine-tune an image model (opens in a new tab)checked
The fal and Replicate pages quoted here are for FluxFlux Flux is a family of image-generation models from Black Forest Labs. It is widely used for photorealistic characters because it holds faces better than earlier open models. Open the glossary. WaveSpeed’s list also has trainers for Z-Image and for Wan 2.2, a video model family. A LoRA only works on the model it was trained for, so pick the host that has a trainer for the model you generate with. The image count is the same everywhere: see how many images to train a LoRA.
Which host runs your persona's main model cheapest right now? Ask it in #technical-inquiries. Someone there has usually had the same problem.
Which host to pick for an AI persona
Pick the model first and the host second. The model decides whether your persona keeps her face. The host only decides what you pay for it.
- Find the model that holds your persona’s face. Which AI video models hold a face shows how to test that.
- Price that one model on all three hosts at your real resolution, length and sound setting. Ignore promotions.
- Work out one month with your real mix of images and clips, as in the example above.
- Check the rules that cost you time: credit expiry, two jobs at once for new accounts, and Replicate’s one-hour file limit.
- Keep one host for most work. A second account only pays off when one model you use often is much cheaper there.
We use WaveSpeed for persona work because one account covers images, video and LoRA training, and its credit does not expire. AI OFM tools shows where a host fits next to the rest of your stack, and the operations guide shows how to keep the weekly work small.
Your next step: write down the one image model and the one video model you use most. Open their pages on WaveSpeed, fal and Replicate today and put the three prices next to each other. If one host is more than 30% cheaper for your main video model, move that model there.
Sources for this article
The platform rules, laws and numbers above come from these documents. The date says when we last opened each one and checked it.
- WaveSpeedAI — Nano Banana 2 Editchecked
- WaveSpeedAI — Nano Banana Pro Editchecked
- WaveSpeedAI — Kling v3.0 Pro Image-to-Videochecked
- WaveSpeedAI — Kling v3.0 Std Image-to-Videochecked
- WaveSpeedAI — Seedance 2.0 Image-to-Videochecked
- WaveSpeedAI — MiniMax H3 Image-to-Videochecked
- fal — Nano Banana 2 Editchecked
- fal — Nano Banana Pro Editchecked
- fal — Kling Video v3 Pro Image-to-Videochecked
- fal — Kling Video v3 Standard Image-to-Videochecked
- fal — Seedance 2.0 Image-to-Videochecked
- fal — MiniMax H3 Image-to-Videochecked
- Replicate — google/nano-banana-2checked
- Replicate — google/nano-banana-prochecked
- Replicate — kwaivgi/kling-v3-videochecked
- Replicate — bytedance/seedance-2.0checked
- Replicate — minimax/h3checked
- WaveSpeedAI — Z-Image Turbochecked
- fal — Z-Image Turbochecked
- Replicate — prunaai/z-image-turbochecked
- fal — MiniMax H3 Max Image-to-Videochecked
- Replicate pricingchecked
- fal docs — Model APIs FAQchecked
- Replicate docs — Prepaid creditchecked
- Replicate docs — Billingchecked
- WaveSpeedAI docs — Refund Policychecked
- WaveSpeedAI docs — How Pricing Workschecked
- WaveSpeedAI docs — Account Levels & Rate Limitschecked
- Replicate docs — Rate limitschecked
- Replicate docs — Data retentionchecked
- WaveSpeedAI docs — Data Retention Policychecked
- WaveSpeedAI — Training Tools collectionchecked
- fal — Flux LoRA Fast Trainingchecked
- Replicate docs — Fine-tune an image modelchecked
Questions people ask
Is WaveSpeed cheaper than fal?
For most of the models we priced. On 5 October 2026 WaveSpeed and fal charged the same for Kling 3.0, WaveSpeed was cheaper for Nano Banana 2, Nano Banana Pro and Seedance 2.0 at 720p, and fal was cheaper for MiniMax H3. Compare the model you use, at your settings.
Is fal cheaper than Replicate?
Not always. On 5 October 2026 fal cost half as much as Replicate for Kling 3.0 and less for MiniMax H3. Replicate was cheaper for Nano Banana 2 and Seedance 2.0 at 720p.
What is the cheapest Kling 3.0 API?
On 5 October 2026 WaveSpeed and fal both charged $0.56 for 5 seconds of Kling 3.0 Pro image-to-video without sound, and $0.42 for Std. Replicate charged twice that. Sound makes every host more expensive.
Do API credits expire on WaveSpeed, fal and Replicate?
WaveSpeed credits never expire. Purchased fal credits expire 365 days after purchase, and Replicate credit is valid for one year. None of the three refunds unused credit.
Which host is best for AI video?
No single host. Each one was cheapest for a different video model on the same day. Pick the model that holds your persona's face first, then check which host sells that model cheapest at your resolution and length.