Summary: Replicate, fal.ai, and Fireworks AI all rent you inference without you racking a GPU, but they invoice three different things: Replicate mostly bills public-model runtime (or a listed input/output price), fal bills Model APIs per successful output from prepaid credits and Serverless GPUs by the second, and Fireworks bills serverless chat by the token with Standard, Priority, Fast, Batch, and Spot modes. This comparison, checked against those official pricing pages and docs on 2026-11-08, walks through how each one charges, how to make a first documented API call, and which setup matches image/video catalogs versus open-model text serving. Official screenshots are at the end of this post.
Three invoices for the same idea
People lump these platforms together because each one lets you call a hosted model with an API key instead of standing up CUDA yourself. The bill, though, is not “GPU time” in the same sense on all three, and mixing the units is how teams overspend or pick the wrong default.
Replicate is a catalog of public and private models — image, video, audio, and language — where most public models charge only while they are actively processing your request, and some charge a published input/output price instead. fal splits into two products on purpose: Model APIs (call 1,000+ hosted image, video, audio, and multimodal endpoints) and fal Serverless (deploy your own Python app on fal GPUs). Fireworks is built around fast serverless text and vision models with per-token rates, plus on-demand GPUs, training, and a separate Nexus coding plan on the same pricing page.
If you already route several providers through one OpenAI-compatible URL, our LiteLLM tutorial is the layer that sits in front of a choice like this. The decision below is which backend you point that layer at.
Replicate: pay for a run, not a monthly seat
Replicate’s pricing page opens with a single rule: you only pay for what you use. Some models are billed by the hardware and the time they take; others are billed by input and output. The page tells you to read the estimate on each model’s own page rather than treat one number as universal.
Public models, private models, and deployments
On a public model, the billing docs say you pay only for the time the model is active and processing your requests. Setup and idle time are free. You share a hardware pool and a queue with other customers, so you can hit cold boots or scaling limits when the shared pool is busy. If you want your own instances and queue, Replicate points you at a deployment.
Private models (with the exception of fast-booting fine-tunes) run on dedicated hardware. You pay for setup, idle time waiting for requests, and active processing. Fast-booting fine-tunes are labeled in the model’s version list; for those, Replicate only charges active processing time, public or private.
Deployments let you pin hardware and scaling for any model, including a public model you do not own. Billing matches private models: setup, idle, and active all count. The docs suggest a deployment when you want different hardware than the public default, a predictable request rate, or insulation from other customers on a shared pool. They also note that a well-tuned deployment is usually only marginally more expensive than the public model, because setup and idle should be a small slice of active time if you size it correctly.
Failed runs are not charged on ordinary models. If you cancel a run for an official model, you may still be charged for the time it already ran. On private models and deployments, failed and canceled runs are billed for the time their instances were active. If a model calls other models, you pay for the root run and for downstream models it invoked; a failure still bills the root duration plus any downstream work that already happened.
Hardware prices and a few listed model rates
These hardware figures are the public list on replicate.com/pricing on 2026-11-08:
- CPU (Small) /
cpu-small: $0.000025/sec ($0.09/hr), 2 GB RAM - CPU /
cpu: $0.000100/sec ($0.36/hr), 8 GB RAM - Nvidia T4 /
gpu-t4: $0.000225/sec ($0.81/hr), 16 GB GPU RAM - Nvidia L40S /
gpu-l40s: $0.000975/sec ($3.51/hr), 48 GB GPU RAM - Nvidia A100 80 GB /
gpu-a100-large: $0.001400/sec ($5.04/hr) - Nvidia H100 /
gpu-h100: $0.001525/sec ($5.49/hr) - Nvidia H200 /
gpu-h200: $0.001525/sec ($5.49/hr); the page says H200 capacity is available with committed-spend contracts
Multi-GPU rows (2x/4x/8x A100, H100, H200, L40S) are listed as multiples of the single-GPU rate. Several of those larger SKUs are flagged as needing committed-spend contracts.
Examples of models that bill by input and output, copied from the same pricing page that day:
black-forest-labs/flux-schnell: $3.00 per thousand output imagesblack-forest-labs/flux-dev: $0.025 per output imageblack-forest-labs/flux-1.1-pro: $0.04 per output imagewavespeedai/wan-2.1-i2v-480p: $0.09 per second of output videowavespeedai/wan-2.1-i2v-720p: $0.25 per second of output videodeepseek-ai/deepseek-r1: listed as $3.75 per million input tokens and $0.01 per thousand output tokens
Free tries, then prepaid credit
Replicate’s billing docs still describe free limits: you can run select models for free, and after a bit you are asked to set up billing. The Try for Free collection is the public list of those models. The collection page says the free use is a limited number of runs, not a forever-free tier, and it does not publish a single run count that applies to every model.
For paying use, Replicate moved new accounts (from 16 July 2025, per their prepaid-credit changelog) onto prepaid credit instead of monthly arrears. The prepaid credit page says you add a payment method, click Add credit, and usage deducts from that balance. Purchased credit is valid for one year from the purchase date and is not refundable. Auto-reload is optional: the documented minimums are a $5 threshold and a $15 reload balance. When the balance hits zero, Replicate prevents new work and shuts down infrastructure it is running for you; a prediction that overruns the balance can be charged to your default payment method at month-end.
A first Replicate call (from the official Python guide)
The Run a model from Python page documents this path. Create a token at replicate.com/account/api-tokens (tokens are 40-character strings that start with r8_), then:
pip install replicate
export REPLICATE_API_TOKEN=r8_....
import replicate
output = replicate.run(
"black-forest-labs/flux-schnell",
input={"prompt": "an iguana on the beach, pointillism"}
)
with open("output.png", "wb") as f:
f.write(output[0].read())
Every model also has a web playground. The docs treat that form as the way to learn the inputs, then move the same call into the API. We did not run this snippet against a live token for this post; the code and token format are what the official pages show.
fal.ai: two products, two meters
fal’s pricing page states the split in one sentence: Model APIs are priced per output, per model, and drawn from prepaid credits; fal Serverless GPUs are priced per hour and billed per second.
Model APIs: pay for a successful output
The Model API pricing docs say you are billed for successfully generated outputs and never for server errors (HTTP 500 or higher) or for time spent waiting in the queue. Client-side errors such as invalid inputs (HTTP 422) may still be charged if a runner already spent GPU time before the error was detected, according to the Model APIs FAQ. Cold starts on Model API endpoints — container pull and model loading — are not charged.
The billing unit is per model: typically per image or per megapixel for image models (higher resolutions cost proportionally more when the unit is megapixels), per second or per video for video models, and per request or another output unit for audio and other models. Models without a fixed per-output price fall back to per-second GPU billing. The exact current price of any model is on that model’s page, on fal.ai/pricing, or via:
curl "https://api.fal.ai/v1/models/pricing?endpoint_id=fal-ai/flux/dev" \
-H "Authorization: Key $FAL_KEY"
A selection from the public pricing page on 2026-11-08 (units copied as printed there):
fal-ai/flux/schnell(FLUX.1 [schnell]): $0.003 per megapixelsfal-ai/flux/dev(FLUX.1 [dev]): $0.025 per megapixelsfal-ai/nano-banana: $0.0398 per imagesfal-ai/nano-banana-2: $0.08 per imagesfal-ai/kling-video/v2.6/pro/image-to-video: $0.07 per secondsfal-ai/kling-video/v3/pro/image-to-video: $0.14 per secondsfal-ai/minimax/speech-02-hd: $0.1 per 1,000 characters
Do not line those FLUX numbers up against Replicate’s per-image FLUX prices without converting units. $0.003 per megapixel is not the same as Replicate’s $3.00 per thousand output images ($0.003 per image) unless the output is actually one megapixel.
Credits are prepaid. Purchased credits expire 365 days from the purchase date. The FAQ says free credits and coupons have variable expiration (from one week to one year) depending on the grant — the public docs do not state a single starter amount, so this post does not invent one. When the balance drops below your account’s lock threshold, the account is locked and API requests are rejected until you add credits (enterprise invoice customers are not auto-locked).
Concurrency is tied to those purchases. New accounts start at 2 concurrent requests; as you buy credits the limit increases automatically up to 40. Extra requests are queued. Generated media files stay on fal’s CDN for at least 7 days by default, and the returned https://v3b.fal.media/... URLs are publicly accessible unless you set an ACL. Each model has its own license: look for a Commercial use or Research only badge on the model page rather than assuming every output is cleared for a product.
A first fal Model API call
The official Quick Start and authentication page document this consumer path. Create a key at fal.ai/dashboard/keys, copy it immediately (you will not see it again), and prefer the API preset unless you need Serverless CLI permissions:
pip install fal-client
export FAL_KEY="your-api-key-here"
import fal_client
result = fal_client.subscribe("fal-ai/flux/schnell", arguments={
"prompt": "a futuristic cityscape at sunset"
})
print(result["images"][0]["url"])
The same Quick Start shows a JavaScript fal.subscribe call and a cURL POST to https://queue.fal.run/fal-ai/flux/schnell with Authorization: Key $FAL_KEY. The auth page’s test request uses https://fal.run/fal-ai/flux/schnell. Do not put that key in browser JavaScript; fal’s FAQ says to proxy from your backend.
fal Serverless: billed while the runner is alive
If you want to deploy your own pipeline, fal Serverless is a separate meter. The public GPU card on 2026-11-08 listed:
- H100 80 GB: $4.50/h list ($0.00125/s), as low as $2.49/h
- H200 141 GB: $6.00/h list, as low as $2.99/h
- B200 192 GB: $7.99/h list, as low as $5.49/h
- B300 288 GB: $12.99/h list, as low as $5.99/h
- GB200 192 GB: $9.99/h list, as low as $5.89/h
- RTX PRO 6000 96 GB: $4.00/h list, as low as $1.99/h
Runners are billed per second while they are alive — the documented states are SETUP, IDLE, RUNNING, DRAINING, and TERMINATING. Waiting to be scheduled and pulling the container image are not billed. A multi-GPU runner is billed as GPU count times duration. Per-second prices on that page are the hourly price divided by 3,600. Docs also state that Serverless deploys need access the fal team approves for each account; you request it at fal.ai/dashboard/serverless-get-started.
Fireworks: tokens first, GPUs when you outgrow shared capacity
Fireworks’ pricing page frames three ways to deploy: Serverless (pay per token, no commitment, usage draws down prepaid credits, “from $0.05 / 1M input tokens”), Reserved Throughput (you reserve dollars per minute with an SLA; usage above the reservation can bill at standard Serverless rates, never a penalty rate), and Fireworks Sovereign (contract, private preview for Enterprise: Fireworks-managed inference on GPUs in your cloud or data center).
Modes on the same endpoint
Every serverless text or vision request bills input tokens, cached input tokens, and output tokens. Embeddings bill input only. You can switch modes per request:
- Standard — the default for most traffic
- Priority — prioritize your traffic during peak load
- Fast — described as 100+ tokens per second for interactive apps
- Ultra-fast — preview for specific models; the page says to talk to them
- Batch — 50% off input and output for jobs that do not need an immediate answer
- Spot — 30% off Standard on best-effort capacity, for select models only; throughput can drop to zero, so the page tells you to retry or fall back to Standard
Prompt caching is automatic: no setup and no write fee. Cached input bills at 50% off by default, with deeper discounts on select models (the highlighted table shows some cached rates well below half, such as DeepSeek V4.1 Flash cached input at $0.006 against $0.30 input). The page also notes a 50% premium for US-only Serverless on select models, waived on Muse Spark Open. Serverless Pricing is the source of truth for which models offer Priority (a dash means it is not available).
Highlighted Standard rates from the public page and docs on 2026-11-08, in USD per 1M tokens as input / cached input / output:
- NVIDIA Nemotron 3.5 Lightning 30B A3B: $0.05 / $0.01 / $0.20
- GLM 5.3 Flash: $0.15 / $0.03 / $0.50
- OpenAI GPT OSS 120B: $0.15 / $0.015 / $0.60
- DeepSeek V4.1 Flash: $0.30 / $0.006 / $1.20
- MiniMax M3: $0.30 / $0.06 / $1.20
- Qwen 3.8 Max: $2.00 / $0.25 / $6.00
- Kimi K3: $3.00 / $0.30 / $15.00
- Ember-1: $3.00 / $0.30 / $15.00
Models not listed individually are priced by size on the docs page: $0.10 / 1M tokens under 4B parameters, $0.20 from 4B to 16B, $0.90 above 16B, $0.50 for MoE up to 56B, and $1.20 for MoE from 56.1B to 176B — applied uniformly to input and output, with no separate cached-input rate. Embeddings: $0.008 / 1M input tokens up to 150M parameters, $0.016 from 150M to 350M, and $0.10 for Qwen3 8B.
On-demand GPUs, spend caps, and Nexus
On-demand serving and the Dedicated Training API share one GPU rate card. There is no charge for on-demand start-up time. Idle deployments scale to zero after one hour by default. Region-restricted deployments are priced at 1.5x. Public hourly rates on 2026-11-08: H100 80 GB $8.00, H200 141 GB $8.00, B200 180 GB $13.00, B300 288 GB $15.00, GB300 288 GB $20.00.
You buy credits and usage draws them down. You can set a monthly spend limit; Auto Reload is separate and buys more credits when the balance runs low without raising that limit. The page states zero data retention by default, and lists SOC 2 Type II, ISO 27001, 27701 and 42001, plus HIPAA with a BAA.
A training promo on the same page (checked 2026-11-08) said Muse Glimmer 30B SFT is free through 30 November in the UI and the Serverless Training API, with no credit card required. That is a training offer, not a general free inference allowance. Nexus is also on the pricing page and is a different product: Pro at $8/month for individual developers, Team at $10 per seat / month (up to 200 seats), and Enterprise as custom terms.
A first Fireworks serverless call
The Serverless Quickstart has you create an API key in the dashboard, export FIREWORKS_API_KEY, and call GLM 5.3 Flash. The OpenAI-compatible path looks like this:
pip install openai
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ.get("FIREWORKS_API_KEY"),
base_url="https://api.fireworks.ai/inference/v1"
)
response = client.chat.completions.create(
model="accounts/fireworks/models/glm-5p3-flash",
messages=[{"role": "user", "content": "Say hello in Spanish"}],
)
print(response.choices[0].message.content)
The same page also documents the fireworks-ai SDK, an Anthropic-compatible base URL at https://api.fireworks.ai/inference, streaming, function calling (example model accounts/fireworks/models/kimi-k3), JSON-schema structured outputs, and a reasoning_effort parameter on models that support reasoning. The documented curl endpoint is https://api.fireworks.ai/inference/v1/chat/completions.
Replicate vs fal.ai vs Fireworks, side by side
Figures below were read from the official pricing and docs pages linked in Sources on 2026-11-08. Model catalogs move; confirm a rate on the model page or pricing API before you ship.
| Replicate | fal.ai | Fireworks AI | |
|---|---|---|---|
| What you are mainly buying | A large catalog of public (and your own private) models across image, video, audio, and language, plus Cog-packaged custom models | Model APIs for 1,000+ hosted image, video, audio, and multimodal endpoints, or fal Serverless to deploy your own app | Serverless open-model text/vision inference billed per token, with on-demand GPUs, training, and a separate Nexus coding plan |
| How the meter works | Most public models: hardware time while the run is active (setup and idle free). Some models: published input/output prices. Private models and deployments: setup + idle + active | Model APIs: per successful output (image, megapixel, video second, and so on) from prepaid credits. Serverless GPUs: per second the runner is alive | Serverless: input, cached input, and output tokens drawn from prepaid credits. On demand: per GPU-second at the published hourly card |
| Starting without a contract | Select models in the Try for Free collection have a limited number of free runs, then you set up billing. New paying accounts purchase prepaid credit | Prepaid credits; new Model API accounts start at 2 concurrent requests (rises with purchases up to 40). No single public “free credit” amount is stated in the docs we checked | Serverless has no commitment. You buy credits and can set a monthly spend cap. Muse Glimmer 30B SFT was listed as free through 2026-11-30 (training only) |
| Example public rates (2026-11-08) | flux-schnell $3.00 / thousand output images; flux-dev $0.025 / image; H100 $5.49/hr ($0.001525/sec) | FLUX.1 [schnell] $0.003 / megapixel; FLUX.1 [dev] $0.025 / megapixel; Kling v2.6 Pro I2V $0.07 / second; H100 Serverless list $4.50/h (as low as $2.49/h) | Nemotron 3.5 Lightning 30B $0.05 / $0.20 per 1M input/output; GLM 5.3 Flash $0.15 / $0.50; Batch 50% off; Spot 30% off Standard; on-demand H100 $8.00/hr |
| When idle or queue time is free | Public models: setup and idle are free. Private models and deployments: you pay both | Model APIs: queue wait and HTTP 500+ errors are free; Serverless does not bill waiting-to-schedule or image pull, but does bill SETUP and IDLE once the runner is alive | On-demand start-up time is not charged; idle deployments scale to zero after one hour by default |
| Credits running out | New work stops and running infrastructure is shut down; a rare overrun can be charged to your card at month-end. Purchased credit lasts 1 year and is not refundable | Account locks below the lock threshold and requests are rejected (except invoice enterprise). Purchased credits expire after 365 days | Usage draws down prepaid credits; monthly spend limits cap the bill, while Auto Reload can buy more credits so production keeps running |
| Best when | You want to try many community and official models, including ones you package with Cog, and you prefer paying for each public run’s active time | Your product is image, video, or audio generation against a hosted catalog, or you want to deploy your own generative pipeline on fal GPUs | You want OpenAI- or Anthropic-shaped chat completions against open models, with batch/spot/cache levers and a path to reserved or on-demand capacity |
Editor's note
Who it's for: developers who need an API for image/video generation or for hosted open-model text, and who want to compare the actual invoice unit before they hard-code a vendor. Teams that already use a gateway such as LiteLLM and need to pick one backend for a workload.
Who it's not for: someone who only wants a consumer chat website with a monthly personal plan. Anyone hoping for a single “cheapest FLUX” number that works on every platform — the units (per image vs per megapixel vs per second of GPU) are not interchangeable. Anyone who needs an SLA on shared public queues without paying for reserved or dedicated capacity.
Gotchas verified on official pages (2026-11-08):
- Replicate public vs private billing is easy to mix up. Public setup/idle is free; private models and deployments charge for both.
- Replicate purchased credit expires in one year and is not refundable. Auto-reload minimums are a $5 threshold and a $15 reload balance.
- fal Model API 422 errors can still cost money if a runner already used GPU time. Only HTTP 500+ is promised free.
- fal output URLs are public by default for at least seven days. Set an ACL or download files if that is a problem.
- fal Serverless is not self-serve for every account. Docs say deploys need approved access.
- Fireworks Spot can drop to zero throughput. The pricing page tells you to retry or fall back to Standard. Batch is the 50% off path for work that can wait.
- Fireworks US-only serverless has a 50% premium on select models (waived on Muse Spark Open). Check the (US) rows on the serverless pricing docs before you assume the headline rate.
- Nexus $8 / $10-per-seat is not serverless inference. It is a separate coding-tool plan on the same Fireworks pricing page.
FAQ
Is any of these three free to run in production?
Not as a standing free production tier. Replicate lets you try select models for a limited number of runs, then asks you to set up billing and (for new accounts) buy prepaid credit. fal and Fireworks both draw production usage from prepaid credits; fal’s public FAQ does not publish one starter credit figure, and Fireworks’ no-credit-card promo we saw on 2026-11-08 applied to Muse Glimmer 30B SFT training through 30 November, not to unlimited serverless chat. You still pay the model provider rates once you are past those documented try or promo paths.
Which one should I use to generate images or short videos?
Start with the catalog that actually lists the model you want, then read that model’s unit. Replicate and fal both publish a wide image and video catalog with per-image, per-megapixel, or per-output-second prices. Fireworks’ public pricing emphasis is serverless text and vision tokens, plus on-demand GPUs; use it when the workload is chat, tools, or embeddings rather than “give me a Kling clip.” Compare fal’s megapixel prices to Replicate’s per-image prices only after you know the output resolution.
Can I put all three behind one OpenAI client?
Fireworks documents an OpenAI-compatible base URL (https://api.fireworks.ai/inference/v1) and an Anthropic-compatible URL, so many OpenAI or Anthropic SDK callers work with a base-URL change. Replicate and fal use their own clients (replicate.run, fal_client.subscribe) and HTTP APIs shaped around predictions or a queue; they are not drop-in Chat Completions endpoints. A gateway such as LiteLLM is how teams hide that difference, and you still have to configure each provider the way its own docs describe.
Related posts on lightmind24
- Together AI Pricing: Credits, Rates and Limits — another hosted inference bill, useful if your workload is tokens on open models rather than a generative catalog.
- Groq API Free Tier: Rate Limits, Costs and When to Upgrade — speed-focused token inference with a documented free tier, a different tradeoff from these three catalogs.
- LiteLLM Tutorial: One Proxy for OpenAI and Ollama — put whichever backend you choose behind one OpenAI-compatible URL.
Pick the platform whose invoice unit matches the work: active seconds or listed output prices on Replicate, per-output credits on fal Model APIs, or per-token serverless modes on Fireworks — then confirm today’s rate on that model’s official page before you ship.
Last verified: 2026-11-08
Sources: Replicate pricing, Replicate billing, Replicate prepaid credit, Replicate Python guide, Replicate Try for Free, Replicate prepaid-credit changelog, fal pricing, fal Model API pricing, fal Model APIs FAQ, fal Quick Start, fal authentication, Fireworks pricing, Fireworks serverless pricing, Fireworks serverless quickstart.
Image credit: screenshots of the public Replicate pricing page, fal.ai pricing page, Fireworks AI pricing page, and Fireworks serverless quickstart docs, captured 2026-11-08. See the official screenshots at the end of this post. Includes no AI-generated image.
댓글
댓글 쓰기