Most API price comparisons are apples to oranges: different models, different quality, different products. FLUX hosting is the rare apples-to-apples case. Black Forest Labs released the FLUX weights, multiple inference platforms host them, and the result is the same model available through three vendors in our catalog at prices spanning 3.3x. This is as close to pure commodity arbitrage as the AI tool market gets, and it is worth understanding both because the savings are real and because the ways the “same model” still differs are instructive.
Here are the three FLUX-dev hosts we route, cheapest first, with routed prices (provider list plus 20%; the pricing docs explain the margin).
1. DeepInfra: $0.0108 per image
DeepInfra is the price floor for FLUX-dev, and honestly for the whole image category: just over a cent per image routed. Our catalog scores it 80. DeepInfra’s business is running open-weight models cheaply at scale, and the pricing reflects a company competing on exactly one axis. If your workload is high-volume generation where each individual image is low-stakes (thumbnails, drafts, variations for a human to pick from), it is very hard to argue with the cheapest FLUX on the market.
2. fal.ai: $0.03 per image
fal.ai hosts flux-dev at $0.03 per image routed and scores highest of the three in our catalog at 84. fal has built its reputation specifically on media model inference, and beyond flux-dev it also serves flux-pro, so it is the host to pick when you want headroom into the stronger FLUX variant without changing vendors. In my experience the platforms that specialize in image and video inference earn their premium on the product surface around the model: the API ergonomics and the model breadth, not the pixels themselves.
3. Replicate: $0.036 per image
Replicate serves FLUX-dev at $0.036 per image routed with a catalog score of 82. Replicate’s catalog of hosted models is enormous, and that is its structural appeal: if your product touches many models, having FLUX where everything else already lives is a real convenience. As a pure FLUX-dev price play, though, it is the most expensive of the three, at 3.3x DeepInfra’s price.
The table
| Host | Routed price per image | Catalog quality score | Notes |
|---|---|---|---|
| DeepInfra | $0.0108 | 80 | Price floor |
| fal.ai | $0.03 | 84 | Also serves flux-pro |
| Replicate | $0.036 | 82 | Broad model catalog |
For scale: at 100,000 images a month, DeepInfra costs $1,080 where Replicate costs $3,600. Same weights. That gap is a junior engineer’s worth of money per year at sustained volume, and capturing it requires changing a URL.
If the model is identical, why do the scores differ?
Fair question, and it gets at what our quality scores actually measure. They are hand-curated 0-100 ratings of the provider experience, not just the model output: reliability, API design, documentation, feature coverage. Two hosts serving identical weights can still differ in how often they error, how they queue requests, and what knobs they expose. The 80 to 84 spread across these three is small precisely because the underlying model is the same; the differences are in the wrapper. Compare that with the search category, where scores span 72 to 90 because the products are genuinely different things.
The different-model tier, for contrast
The FLUX trio is the commodity end of our image catalog. The other end is proprietary models where no arbitrage exists because there is exactly one host:
- OpenAI GPT Image: $0.0636 per image routed, catalog score 90, the quality ceiling of the category.
- xAI Grok: $0.084 per image routed, score 83.
The pricing logic flips here. You are not shopping between hosts of one model; you are deciding whether a different, stronger model is worth roughly 2x to 8x the FLUX-dev floor. That is a quality decision, not an arbitrage. Both kinds of decision live side by side on the image generation comparison page.
What arbitrage means for routing
Same-model hosting is the ideal case for cheapest-first routing with failover, because the usual objection (“the cheap provider might be worse”) mostly evaporates when the weights are identical. My default for FLUX-dev traffic: route cheapest-first, so DeepInfra takes the volume, with fal.ai and Replicate as the failover chain behind it (the router tries up to 3 providers per request). If DeepInfra has a bad five minutes, the circuit breaker shifts traffic to fal automatically, and every response carries the routing.attempted chain so you can see exactly which host produced which image. I covered the general failover mechanics in why AI agents need failover.
The broader lesson generalizes past FLUX. Whenever open weights get popular, a hosting market forms around them, prices diverge, and the spread persists longer than you would expect. Whisper hosting shows the same pattern in transcription. Checking whether your model has a cheaper host is the fastest cost win in the agent stack.
Current prices for all five image providers, from the same open catalog our router reads, are on the image generation API comparison page.