Marketplaces like fal and Replicate are now core infrastructure for image generation. Here is when they are the right door, when to go direct, and the hybrid most production teams end up with.
.png)
Recraft's models are built for aesthetics first: output that looks designed rather than merely generated, with tight prompt adherence and controls for style, color, and format. That matters for this article because the way you call a model decides which of those controls you actually get.
This week fal closed a $1.3 billion round at a $16 billion valuation, with reported annualized revenue above $800 million. An AI model marketplace is no longer a convenience layer for hobby projects. It is infrastructure, and a real share of production image generation now runs through fal, Replicate, and similar platforms rather than through the model labs themselves. If you are choosing where to call an image model from, the question is not "marketplace or direct" in the abstract. It is which door fits your volume, your feature needs, and your tolerance for a middle layer. Here is how to decide.
A marketplace such as fal.ai or Replicate hosts many models behind one account, one billing relationship, and one API shape. You sign up once, pick an endpoint, and call it. The platform handles the GPUs, the queueing, the retries, and the per-request metering.
The value is aggregation. A team testing six image models in a week does not want six vendor contracts and six authentication schemes. A product that routes between models by task (a fast model for thumbnails, a stronger one for hero images) wants a single integration surface. Marketplaces also tend to ship new endpoints within hours of a lab's release, sometimes before the lab's own blog post is live, which makes them the easiest place to try something new.
What a marketplace does not do is own the model. It calls the provider's weights or the provider's own API under a commercial agreement, then resells access. That has three consequences you should understand before you build on it.
Marketplaces expose a model through their own parameter schema. For simple text-to-image calls this is fine. For anything provider-specific, coverage varies by endpoint and by how recently it was updated.
Recraft is a useful example because its differentiating controls are exactly the kind that get flattened. On the direct API, custom styles work in two ways: create a style once from one to ten reference images and reuse its style_id on every request, or attach style references to the request itself. There is a match parameter that sets how strictly V4 follows the reference. Native vector output is a separate model line that returns real SVG, not a raster that someone traced afterward.
On a marketplace endpoint you may get a subset. At the time of writing, the fal documentation for its Recraft V3 endpoint lists a base style choice, a size, and a color palette as inputs. That covers many use cases. It does not cover a saved brand style applied across ten thousand assets, which is the use case most brand and marketing teams end up with. Always read the endpoint schema, not the model name, before assuming parity.
The intuitive assumption is that a marketplace charges a markup. For Recraft V4 that is currently not true: fal lists $0.04 per image, and the direct API lists $0.04 per raster image and $0.08 per vector image. Price parity at list is common because labs set the resale terms.
The differences show up at the edges. Direct API units at Recraft are prepaid at $1 per 1,000 units and do not expire, which suits bursty workloads. Marketplaces bill per call on their own terms and can reprice or retire an endpoint on their own schedule. And the moment you need volume pricing, a contract, or an SLA, you are negotiating with the model provider either way. The marketplace cannot sign for the lab.
Every request passes through the marketplace before it reaches the model. That adds latency, usually modest, and adds a second party to your data-handling story, your uptime math, and your incident response. It also adds a strategic question that did not exist two years ago. fal now trains and ships its own post-trained models alongside the ones it hosts. A platform that is both your distributor and a model vendor has incentives that will not always align with yours or with the labs you rely on. That is not a reason to avoid it. It is a reason to know which door you are standing in.
Choose a marketplace when at least two of these are true:
In this situation the marketplace is doing real work for you. Paying list price for aggregation is a fair trade.
Go direct when any one of these is true:
Going direct costs you the integration work the marketplace was absorbing. For one or two core models that is a few days. For ten it is a platform team.
The pattern we see most often is not either-or. Teams prototype on a marketplace because it is the fastest way to compare outputs across labs. Once a model becomes load-bearing for a product, they move that one model to a direct integration, keep the marketplace for experiments and long-tail models, and put a thin routing layer in front of both.
A practical way to decide per model: if you are calling it with only a prompt and a size, leave it on the marketplace. If you are calling it with a style_id, a palette, a vector flag, or anything that encodes your brand, go direct. The features that make the output yours are the ones most likely to be missing in the middle layer.
Recraft's API is available both ways. The marketplace endpoints are the right place to see whether the aesthetic fits your product. The direct API is where custom styles, vector generation, and volume terms live. If you already use an agent workflow, the Recraft MCP server is a third door that sits closer to the direct side.
For a broader comparison of what the leading image APIs offer and cost, see our guide to the best AI image generation APIs in 2026.
An AI model marketplace is a platform such as fal.ai or Replicate that hosts many models behind one account, one billing relationship, and one API shape. It runs the GPUs, queueing, and metering, and resells access to models it does not own. You trade some feature coverage and a second party in your request path for a single integration across many labs.
Usually not at list price. For Recraft V4, fal charges $0.04 per image and the direct API charges $0.04 per raster image, so there is no markup on the base call. The differences appear at volume, where committed-use pricing, contracts, and SLAs come from the model provider, and in billing terms such as prepaid units that do not expire.
Go direct when you depend on provider-specific features the marketplace schema does not expose, such as saved custom styles, reference-image controls, or native SVG output. Also go direct when your volume justifies a contract, when compliance needs a short data path, or when you want a roadmap relationship with the lab. For prompt-and-size calls at low volume, a marketplace is the simpler choice.
Check the specific endpoint schema before assuming so. At the time of writing, marketplace endpoints for Recraft document base style presets, sizes, and color palettes, while the direct Recraft API supports custom styles created from one to ten reference images, reusable through a style_id, plus a match parameter that controls how strictly V4 follows the reference. If brand consistency is the goal, the direct API is the reliable path.
.svg)
