fal vs Replicate vs Ofox: Video API Pricing (2026)

TL;DR: Three video APIs, three billing models, one question: who is cheapest for real volume. The answer only means something at a matched model and resolution. On Seedance 2.0 at 1080p, Ofox bills $0.34/s ($1.70 per 5-second clip) against fal’s $0.682/s with audio ($3.41), so Ofox is about half; on Wan 2.6 and 2.7 the two are identical ($0.10/s at 720p, $0.15/s at 1080p). fal only wins on a lower tier, like Wan 2.5 at $0.05/s (480p) or the prior-generation Seedance 1 Pro, and Replicate’s official per-second video sits higher (Wan 2.1 720p at $0.25/s). Every Ofox rate is a 480p “from” floor that steps up with resolution. A per-second number without a model version and a resolution attached is not a price you can act on.

Which One Should You Pick?

If you run the newest Seedance or Wan at scale, Ofox is the cheapest anchor on a matched model and resolution: the same Seedance 2.0 at 1080p costs about half what fal charges, and Wan 2.6/2.7 costs exactly the same on both. fal wins only on a model it prices low at a lower tier (Wan 2.5 at $0.05/s, 480p) or a first-party model Ofox does not host (Kling, Veo, Sora). Replicate’s official per-second video models run higher, and its hardware-per-second track is the odd one out: powerful for custom or open-weight models you deploy yourself, but hard to forecast because you pay for render time, not output seconds. It also lists an older Seedance lite line cheaper than anyone, at a prior generation’s quality.

Everything below is the arithmetic behind that table: each platform’s real pricing pulled from its own page this week, a 2,000-clip monthly bill run three ways, and the gotchas that make a “cheap” per-second rate lie to you.

Quick Specs Comparison

The three platforms do not price video the same way, which is the whole reason a naive “per second” comparison misleads people. Ofox is per output second, stepped by resolution. fal is per-model, sometimes per second, sometimes per whole video, sometimes per token. Replicate has two entirely separate billing tracks. Here is the shape of each, with numbers pulled from each vendor’s live page on 2026-07-20.

Two things to read off this. First, on a matched model and resolution Ofox is the cheapest current-generation option: the same Seedance 2.0 at 1080p is $0.34/s against fal’s $0.682/s, and Wan 2.6/2.7 costs the same on both. Second, fal’s low headline numbers buy a different tier, not a straight discount: Wan 2.5 at $0.05/s is 480p, and Seedance 1 Pro is the prior generation. Replicate lists the single cheapest raw per-second rate here, an older Seedance lite line near $0.018/s, but its current official video models (Wan 2.1 720p at $0.25/s) sit above the others, and its hardware track is a different animal entirely, covered below.

One thing Ofox does not do: it does not host Kling or Hailuo. If your pipeline is committed to Kling specifically, fal is where you get it at $0.07/s. This piece is about price for comparable capability, not a claim that Ofox has every model.

How Each Platform Actually Bills

Before any monthly-bill math, you have to understand that “per second” means three different things across these platforms.

Ofox is the simplest to reason about, but its rate is not flat across resolutions. Every video model bills per output second, and the number on the catalog is the model’s lowest “from” price, its 480p floor. bytedance/seedance-2.0 is $0.07/s at 480p, $0.16/s at 720p, and $0.34/s at 1080p, so a 5-second 1080p clip is 5 × $0.34 = $1.70. There is no hardware line and no idle charge, but you have to read the tier for your resolution, not the headline floor. You submit to POST /v1/videos, poll until the task is completed, and the bill is duration × the rate for that resolution. Video-to-video runs can bill higher than text-to-video on the same model.

fal bills per model, and the unit changes model to model. Its pricing page states billing is “by output unit, per second or per video, depending on the model.” Some models are clean per-second: Wan 2.5 at $0.05/s (480p), Wan 2.6/2.7 at $0.10–$0.15/s, Kling 2.5 Turbo Pro at $0.07/s, Veo 3 at $0.50/s audio-off or $0.75/s audio-on (Veo 3 Fast is $0.25–$0.40/s). Others are per whole video, like Ovi at $0.20/video. Seedance is priced per million video tokens: fal’s current Seedance 2.0 works out to $0.3034/s at 720p and $0.682/s at 1080p with audio, while the older Seedance 1 Pro lands near $0.124/s. You cannot assume a fal model is per-second, or which generation the listing is, until you read its page.

Replicate has two tracks that catch people out. Most public models are billed by hardware time, meaning you pay for the seconds the GPU runs, at rates from $0.000025/s for a small CPU (a plain CPU is $0.0001/s) up to $0.001525/s for an H100. A video that takes 90 seconds to render on an H100 costs 90 × $0.001525 = $0.137 in GPU time, but render time is variable and hard to predict before you run. Separately, Replicate’s curated official models bill per output second like the others: wan-2.1-i2v-720p is $0.25/s, wan-2.1-i2v-480p is $0.09/s, and its official Seedance line runs per second too, from an older lite tier as low as $0.018/s up through current models whose rates are listed on each model’s page. For public models you pay only active processing time; private models and deployments bill for setup and idle time too.

The same 1080p clip is a fixed $1.70 on Ofox, a per-model amount on fal (Seedance 2.0 at $3.41 with audio), and either GPU render time or a per-second official rate on Replicate.

Comparing the Same Model: Seedance

The only honest way to compare pricing is to fix everything except the platform. Both fal and Ofox now host Seedance 2.0, so it is the cleanest same-generation anchor. Take a 1080p, 5-second clip with audio.

Ofox runs bytedance/seedance-2.0 at $0.34/s for 1080p, so 5 × $0.34 = $1.70 per clip. fal hosts the same Seedance 2.0 generation, priced per video token, which works out to $0.682/s at 1080p with audio, or about 5 × $0.682 = $3.41 per clip. At 720p the pattern holds: Ofox is $0.16/s ($0.80) against fal’s $0.3034/s ($1.52). On matched Seedance 2.0 at the same resolution, Ofox is roughly half the price, $1.70 versus $3.41 at 1080p, about 47% cheaper at 720p.

One honest caveat, because a knowledgeable reader will raise it if you do not: fal still lists the prior-generation Seedance 1 Pro at about $0.124/s (~$0.62 per 1080p 5-second clip), which is cheaper than either platform’s 2.0. If that older model’s quality clears your bar, it is the lowest 1080p Seedance rate in this comparison. The half-price claim is specifically same-generation, Seedance 2.0 against Seedance 2.0. Replicate also hosts official Seedance per second, from an older lite tier near $0.018/s; its current Seedance 2.0 rate lives on the model page and is not pinned here.

Comparing the Same Model: Wan

Wan is the tie, and it is worth being precise about it. fal and Ofox both host Wan 2.6 and 2.7 at the same rate: $0.10/s at 720p and $0.15/s at 1080p. Price is not the differentiator here; the differentiator is that Ofox runs Wan behind the same single endpoint and schema as Seedance, while on fal each model carries its own schema. fal also lists the older Wan 2.5 with a $0.05/s floor, but that floor is 480p; Wan 2.5 is $0.10/s at 720p and $0.15/s at 1080p, the same as the newer versions. Replicate’s official wan-2.1-i2v-720p is $0.25/s and wan-2.1-i2v-480p is $0.09/s, an older Wan generation priced above both.

The one genuinely cheaper number is fal’s Wan 2.5 at $0.05/s, and it buys 480p. At 720p and above, fal and Ofox charge the same for Wan, so the choice comes down to whether you want Wan behind one endpoint with Seedance (Ofox) or in fal’s per-model catalog. The lesson repeats: a per-second number without a version and a resolution attached is not a price you can act on.

Pricing Math: Real Monthly Bill

Per-second rates are abstract until you multiply by real volume. Take a small content team generating 2,000 clips per month, 5 seconds each, and price each option at the resolution that model is actually used at. That is a realistic cadence for an e-commerce or social team churning product and promo variants. Here is the same workload across all three platforms, each per-clip figure derived from the rate.

Same 2,000 clips a month, each at the resolution its model is used at. Matched Seedance 2.0 at 1080p is $3,400 on Ofox versus $6,820 on fal; Wan 2.6/2.7 ties at $1,500. The cheap floors ($500, $800) are 480p or Mini tiers.

Read the spread by matching models, not by scanning for the smallest number. The two Seedance 2.0 rows are the like-for-like comparison: at 1080p, Ofox is $3,400/month against fal’s $6,820, almost exactly half, and at 720p it is $1,600 against $3,034. Wan 2.6/2.7 is a dead tie at $1,500 on either platform. The genuinely cheap outliers are a different tier: fal’s Wan 2.5 at $500 is 480p, and fal’s older Seedance 1 Pro at $1,240 is a generation back. Replicate’s official Wan 2.1 720p line at $2,500 is the priciest current per-second option shown, though its older Seedance lite tier (~$0.018/s, off this chart) is cheaper than anything if that quality clears your bar.

Two honest asterisks. First, Replicate’s hardware track is not on this chart because it bills render time, not output seconds, so a fast open-weight model on a cheap GPU could come in below any of these, at the cost of predictability. Second, the cheapest raw rates here belong to older or lower-resolution models (fal Wan 2.5 at 480p, Replicate’s Seedance lite), not to a current-generation 1080p clip. The generalizable takeaway is not “Ofox always wins,” it is “on the same current Seedance 2.0 at the same resolution, Ofox is about half fal’s price, Wan is a tie, and nobody surprises you with idle or hardware charges except Replicate’s private track.”

Resolution and Duration: The Hidden Multiplier

A per-second rate is only half a price. The other half is how many seconds and at what resolution, and both platforms and models hide multipliers in those two dimensions that turn a “cheap” rate into a real bill.

Duration is the linear one. Every model here bills per second, so the gap between two rates scales directly with clip length. On matched Seedance 2.0 at 1080p, Ofox at $0.34/s versus fal’s $0.682/s is a $0.342/s spread. On a 5-second clip that is $1.71; on a 15-second clip, the longest either supports, it is $5.13 per clip. Generate a few thousand long clips a month and that spread becomes a line item finance notices. Short clips hide the model choice; long clips expose it. If your format runs long, test and price at your real duration, not a safe 5-second sample, because the cheaper-looking option at 5 seconds is not necessarily the cheaper one at 15.

Resolution is the multiplier people forget, and on Ofox it is not flat. The per-second rate steps with resolution inside the same model: Seedance 2.0 goes $0.07 (480p) → $0.16 (720p) → $0.34 (1080p) → $1.37 (4K), so dropping a resolution drops the bill directly, and the catalog’s “from” price is always the 480p floor. You can save either by dropping resolution on the flagship or by dropping to a cheaper tier (Mini at $0.04/s from, capped at 720p). fal’s token-priced models scale the same way by construction: Seedance computes tokens as height × width × FPS × duration / 1024, so a 1080p clip costs more tokens than a 720p clip of the same length. Replicate splits it across separate model listings, where the 480p Wan model is $0.09/s and the 720p Wan model is $0.25/s, so on Replicate resolution is a model choice, not a parameter.

The practical rule holds across all three: pick resolution by your distribution ceiling, not by the highest number on the spec sheet. Most social distribution caps at 1080p, since Instagram, TikTok, and YouTube Shorts downscale anything higher, so paying for 4K to feed a vertical story is money set on fire. Where 4K earns its cost is masters that get re-edited, VFX plates that get cropped, and deliverables a client archives at full resolution. Match the resolution to where the clip actually plays, then pick the cheapest tier or model that clears it, and the per-second rate finally becomes a price you can trust.

Credits, Minimums, and Hidden Costs

The rate card is the number vendors advertise. The invoice includes things the rate card does not, and those are where two “similar” prices diverge. Four categories matter for video specifically.

Start with idle and setup billing, the biggest silent multiplier. Ofox and both platforms’ public models bill only for what you generate, with no idle charge. Replicate’s private models, deployments, and always-on endpoints bill for setup and idle time, not just active render seconds. If you stand up a private deployment to cut cold-start latency, you pay for every second it sits waiting for the next request, which can dwarf the per-clip cost on a low-traffic pipeline. Ofox has no equivalent line because there is no instance to keep warm: you submit, it renders, you pay for the output. Read the fine print on any provider before you assume “pay per use” means “pay only when generating.”

Add-on features are the next line to check. Audio, lip-sync, and upscaling are sometimes bundled and sometimes billed separately. On the Ofox Seedance and Wan tiers, synced audio is included in the per-second rate. On fal and Replicate, audio behavior is per-model: fal’s Seedance 2.0 rates above are the audio-on figures, and Veo 3 explicitly charges more with audio on ($0.75/s) than off ($0.50/s), so audio is a rate multiplier on those models, not a freebie. If your pipeline needs audio, price the audio-on rate, not the headline number, and if a model offers lip-sync as a separate endpoint, add that call’s cost to the total.

Free credits deserve skepticism. All three offer some form of starter credit to try the API, and all three change those terms regularly. Promo credits typically expire, and a free-tier ceiling on video is small because video is expensive to render. Treat free credits as a way to validate output quality, not as a cost line in a real budget, and confirm the current terms on each provider’s own page the week you sign up rather than trusting a number from a comparison post, including this one.

Finally, retries and failures. A failed generation still consumed compute somewhere, and how that is billed differs. On output-second billing there is no completed output to bill, so a failed task generally costs nothing: on Ofox, per its docs, usage is only recorded on completed tasks, and fal’s per-second models behave the same way. Replicate is often assumed to be the exception, but its billing docs say plainly, “For all models, if a run fails, we don’t charge you,” so a failed run on a public or official model is free there too. The real exception is Replicate’s private models and deployments, which bill for the time their instances were active regardless of outcome (and canceling an official-model run may still incur a charge). If your prompts push the edge of what a model handles, output-second billing and Replicate’s public models both insulate you from paying for the misses; only an always-on private deployment keeps billing through a failure.

The generalizable point: the platform with the fewest billing dimensions is the easiest to forecast and the hardest to surprise you. Ofox has one dimension, output seconds by resolution. fal has two, the per-model unit and any add-on. Replicate has the most, spanning hardware seconds, output seconds, idle, and setup, which buys flexibility at the cost of predictability.

The Async Workflow on Ofox

All three platforms generate video asynchronously, since a clip takes seconds to minutes to render and nobody holds an HTTP connection open that long. On Ofox the flow is one endpoint. POST /v1/videos returns 202 Accepted with a polling_url, the clip renders in the background, and you either poll GET /v1/videos/{id} until status is completed, or register a callback_url webhook and get called back.

Because Ofox bills per output second, the render taking 30 seconds or 3 minutes does not change your invoice. That is the practical difference from Replicate’s hardware track, where a slow render on an expensive GPU costs more, and the reason forecasting is easier when you pay for output rather than compute.

Try All Three via Ofox: One Key, One Loop

The point of routing Seedance and Wan through Ofox is that swapping models is a one-string change on one endpoint, so you can A/B them without standing up two SDKs. Model IDs used below are the literal strings: bytedance/seedance-2.0 and alibaba/wan-2.7.

Want to run the numbers yourself before committing a pipeline? Ofox bills per output second with one key across every video model, so a Seedance-vs-Wan A/B costs the same to set up as running one.

Python, submit both models on one endpoint

import os, time, requests

OFOX = "https://api.ofox.ai/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['OFOX_API_KEY']}"}

def generate(model, prompt, seconds=5, resolution="1080p"):
    r = requests.post(f"{OFOX}/videos", headers=HEADERS, json={
        "model": model,
        "prompt": prompt,
        "duration": seconds,
        "resolution": resolution,
        "aspect_ratio": "16:9",
    })
    r.raise_for_status()
    task = r.json()
    poll = task["polling_url"]
    while True:
        s = requests.get(poll, headers=HEADERS).json()
        if s["status"] in ("completed", "failed", "cancelled", "expired"):
            return s
        time.sleep(2)  # poll every ~2s

prompt = "A ceramic cup on a wooden table, morning light, steam rising."
for model in ("bytedance/seedance-2.0", "alibaba/wan-2.7"):
    out = generate(model, prompt)
    print(model, out["status"])  # cost is duration x per-second rate for that resolution

Node, same shape

const OFOX = "https://api.ofox.ai/v1";
const H = { Authorization: `Bearer ${process.env.OFOX_API_KEY}`, "Content-Type": "application/json" };

async function generate(model, prompt, seconds = 5, resolution = "1080p") {
  const res = await fetch(`${OFOX}/videos`, {
    method: "POST",
    headers: H,
    body: JSON.stringify({ model, prompt, duration: seconds, resolution, aspect_ratio: "16:9" }),
  });
  const task = await res.json();
  while (true) {
    const s = await (await fetch(task.polling_url, { headers: H })).json();
    if (["completed", "failed", "cancelled", "expired"].includes(s.status)) return s;
    await new Promise((r) => setTimeout(r, 2000)); // ~2s between polls
  }
}

const prompt = "A ceramic cup on a wooden table, morning light, steam rising.";
for (const model of ["bytedance/seedance-2.0", "alibaba/wan-2.7"]) {
  const out = await generate(model, prompt);
  console.log(model, out.status); // one-string swap, same billing shape
}

The only line that changes between the two runs is the model string. On fal or Replicate the same A/B means two different request schemas and, on Replicate, possibly two different billing tracks, so the “swap one string” property is specific to running both models behind one Ofox endpoint.

When to Pick fal

Pick fal when a lower tier it prices cheaply is the model you actually want, or when you need a first-party model Ofox does not host. Wan 2.5 at $0.05/s is the cheapest Wan-family rate in this comparison, so if 480p Wan 2.5 meets your bar, fal wins that job outright. fal is also the place for Kling 2.5 Turbo Pro at $0.07/s, Veo 3, and Sora, which Ofox does not carry. On the models both host, though, fal is not cheaper: Wan 2.6/2.7 is identical, and fal’s Seedance 2.0 is about double Ofox at a matched resolution.

The catch is that fal’s billing unit is per model, not uniform. Some models are per second, some per whole video, some per million tokens. You have to read each model’s page before you can forecast a bill, and a token-priced model like Seedance needs the worked-example math to convert into a per-clip number. Budget for the reading time, not just the render time.

When to Pick Replicate

Pick Replicate when you want to run a custom or open-weight model you deploy yourself, and you want control over the exact hardware. The hardware-per-second model (H100 at $0.001525/s down to a small CPU at $0.000025/s) is the right tool when you have a fine-tune or an obscure open model that no gateway hosts as a managed endpoint. For public models you pay only for active processing, which keeps casual usage cheap.

The catch is predictability and, on official models, price. Hardware billing means your cost depends on render time, which varies run to run, so you cannot quote a fixed per-clip number the way you can with output-second billing. Replicate’s curated official video models do bill per output second, but the current Wan line sits above the others here (720p Wan at $0.25/s); it also lists an older Seedance lite tier near $0.018/s that undercuts everyone at a prior generation’s quality. If you want a managed, forecastable per-clip cost on a current Seedance or Wan, this is the platform that makes that hardest.

When NOT to Pick Any of the Three (and What Else Matters)

If your clips are 2 to 3 seconds, Ofox’s Seedance floor of 4 seconds rules Seedance out, so match the model’s minimum duration to your format before you optimize the rate. If you are locked into Kling specifically, Ofox does not host it, so fal is your per-second Kling source. And if your entire pipeline is one obscure open-weight model, a raw GPU host or Replicate’s hardware track may beat all the managed per-second options on cost, at the price of doing your own ops.

For everyone else generating current-generation Seedance or Wan at volume, the decision is mostly about how predictable you need the bill to be. Output-second billing on Ofox gives you a number you can put in a spreadsheet before you run. That is the differentiator, more than any single row in the price table.

For deeper dives on individual models and the broader landscape, see the Seedance 2.0 vs Wan video API comparison, the Seedance 2.0 Fast vs Mini tier comparison, the roundup of Sora, Veo, and Kling video APIs, the Kling 2.6 Pro video API guide, and the Veo 3.1 Google video API tutorial. For non-video model rates, the AI API pricing comparison covers text and multimodal models.

FAQ

The FAQ section below answers the questions that show up most in People Also Ask for video API pricing. Full answers are in the page’s structured data.

Sources Checked for This Refresh

• fal.ai pricing page: https://fal.ai/pricing
• fal.ai Seedance 2.0 model page: https://fal.ai/models/bytedance/seedance-2.0/text-to-video
• fal.ai Seedance 1 Pro model page: https://fal.ai/models/fal-ai/bytedance/seedance/v1/pro/text-to-video
• fal.ai Wan 2.6 / 2.7 model pages: https://fal.ai/models/wan/v2.6/text-to-video, https://fal.ai/models/fal-ai/wan/v2.7/text-to-video
• fal.ai Kling 2.5 Turbo Pro model page: https://fal.ai/models/fal-ai/kling-video/v2.5-turbo/pro/image-to-video
• Replicate pricing page: https://replicate.com/pricing
• Replicate billing docs: https://replicate.com/docs/topics/billing
• Replicate Wan 2.1 i2v 720p model: https://replicate.com/wavespeedai/wan-2.1-i2v-720p
• Replicate official ByteDance / Seedance models: https://replicate.com/bytedance
• Ofox Seedance 2.0 model page
• Ofox Wan 2.7 model page
• Ofox videos API docs: https://ofox.ai/docs/api/videos
• ByteDance Seed’s Seedance page
• Wan project site