Image generation has the widest price spread of any category I route, and unlike search, half of the spread is for the exact same product. The cheapest image in our catalog costs $0.0108 routed. The most expensive costs $0.084. That is 7.8x, and three of the five hosts in between are serving the same FLUX weights.
All prices below are routed prices (provider list plus 20%, per /docs/pricing), per image, with our hand-curated quality scores.
1. DeepInfra (FLUX-dev): $0.0108
The price floor of the category. DeepInfra hosts FLUX-dev at a price that makes image generation essentially a rounding error: a thousand images for about eleven dollars routed. Quality score 80, which reflects the model more than the host. If your workload is volume (thumbnails, variations, drafts, dataset generation), this is where cheapest-first routing sends it, and it is hard to argue with.
2. fal.ai: $0.03
fal hosts both flux-dev and flux-pro, and scores 84 in our catalog, the highest of the FLUX hosts. You pay about 2.8x DeepInfra’s price for the same dev weights, which sounds bad until you remember we are talking about three cents. For workloads where the host’s serving stack and model selection matter, fal is my default answer in the FLUX tier.
3. Replicate: $0.036
Replicate at $0.036 and quality score 82 sits right next to fal. The catalog gap between them is one score point and half a cent per image. In practice this is the textbook failover pair: same model family, different infrastructure, so an outage at one is unlikely to correlate with an outage at the other. Our router tries up to three providers per request, and fal-then-Replicate is a chain I see resolve cleanly.
4. OpenAI GPT Image: $0.0636
Here the comparison changes character, because GPT Image is a different product, not the same weights on different metal. It scores 90 in our catalog, the highest in the image category, and it costs $0.0636 routed. What you are buying is a different tier: stronger prompt following and text rendering, in my experience, and the kind of output you can put in front of a customer without cherry-picking. At roughly 6x the DeepInfra price, it is expensive per image and still cheap in absolute terms.
5. xAI Grok: $0.084
The ceiling of the category at $0.084 routed, quality score 83. That positioning is awkward: priced above GPT Image while scoring below it, and priced at nearly 8x the FLUX floor while scoring in the FLUX band. I wrote a longer take in the Grok image API guide, but the short version is that Grok’s value in a routing table is diversity (a completely independent upstream), not price-performance.
Same-model arbitrage vs different-model tiers
The useful way to read this list is as two separate markets.
The first market is FLUX hosting: DeepInfra, fal, and Replicate serving overlapping weights at $0.0108 to $0.036. This is true commodity arbitrage. The model is the same; you are shopping for infrastructure, and a 3.3x spread for near-identical output is the kind of inefficiency worth exploiting. Routing cheapest-first inside this tier is close to free money, which is why I covered it in detail in the FLUX hosts comparison.
The second market is model tiers: GPT Image and Grok are different products at different quality points. Arbitrage logic does not apply here. Paying $0.0636 for GPT Image instead of $0.0108 for FLUX-dev is not overpaying; it is buying a different thing. The mistake I see is people picking one tier for all traffic: paying GPT Image prices for thumbnail variations, or shipping FLUX drafts where a hero image needed the top tier.
The math that matters
Per-image prices feel like noise until you multiply. At 100,000 images a month:
| Host | Monthly cost (routed) |
|---|---|
| DeepInfra FLUX-dev | $1,080 |
| fal.ai | $3,000 |
| Replicate | $3,600 |
| OpenAI GPT Image | $6,360 |
| xAI Grok | $8,400 |
A $7,300 monthly gap between floor and ceiling, for a category where most requests genuinely do not need the ceiling. My rule of thumb: route bulk generation to the FLUX tier cheapest-first, reserve GPT Image for the requests where quality is the product, and let failover handle the rest. Every response from our router includes the exact price you paid and the full attempted chain, so you can audit whether your traffic actually matches that split.
Side-by-side pricing for all five hosts, generated from the open-source catalog, lives on the image generation comparison page.