AI Model Marketplace or Direct API: Which Should You Use?

Short answerUse an AI model marketplace such as fal, Replicate, or OpenRouter when you are comparing models or only need a prompt, a size, and a palette. Go direct to the model provider when you need its newest model with your own styles, its full tool set, or volume terms. List prices are usually identical; what differs is which controls each endpoint exposes.

Key takeaways

• Recraft V4.1 costs $0.035 per image on the direct API, on fal, and on OpenRouter, so the marketplace adds no markup at list price.
• The endpoint, not the model name, decides your controls: on fal and OpenRouter, custom styles run on Recraft V4 Styles endpoints, while the V4.1 endpoints accept brand colors but no styles.
• OpenRouter's Recraft V4 Styles models have no reusable style ID, so the $0.005 style fee applies to every styled request instead of once per style.
• A marketplace is the right choice when you are evaluating models, routing between them, or running low volume on base parameters.
• Most production teams run both: a marketplace for evaluation and long-tail models, and a direct integration for the model their product depends on.

What an AI model marketplace does

An AI model marketplace hosts many models from different labs behind one account, one billing relationship, and one API shape. For image generation, fal, Replicate, OpenRouter, and Vercel AI Gateway all work this way. You sign up once, pick an endpoint, and call it, while the platform handles the GPUs, queueing, retries, and 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 by task (a fast model for thumbnails, a stronger one for hero visuals) wants one integration surface. Marketplaces also tend to list new models on the day they launch, which makes them the easiest place to try something new.

This layer is now serious infrastructure. On November 3, 2026, fal raised $1.3 billion at a $16 billion valuation in a Series F led by Bessemer Venture Partners, and a person familiar with the company told Bloomberg that its annualized revenue passed $800 million earlier in the year. A real share of production image generation now runs through these platforms rather than through the labs' own APIs.

What a marketplace does not do is own the model. It runs the provider's model or calls the provider's API under a commercial agreement, then resells access through its own parameter schema. That is why the same model can behave differently depending on where you call it.

Recraft V4.1 on a marketplace vs. the direct API

Recraft makes a useful test case because its distinguishing controls (custom styles, brand palettes, and native SVG output) are exactly the kind a marketplace schema can flatten. We checked each endpoint's published schema and price on November 8, 2026.

Price is identical at list, brand colors work everywhere, and both marketplaces carry Recraft's V4 Styles models, including styled SVG. The gaps are narrower and more specific: neither marketplace applies a custom style to V4.1, and OpenRouter has no reusable style ID.

OpenRouter lists Recraft itself as the provider behind its Recraft endpoints, so it acts as a router in front of the direct API. Replicate also hosts the V4.1 family, the SVG variants, and V4 Styles, so check its model pages for current inputs and pricing.

Three things the marketplace layer changes

1. The endpoint decides which features you get

A marketplace exposes each model through its own parameter schema, and each endpoint is maintained separately. For a plain text-to-image call, that is fine. For provider-specific controls, coverage depends on the endpoint and on how recently someone updated it.

On the direct API, you create a custom style once from up to 10 reference images and reuse its style_id on any V4 or V4.1 request, or attach style references to a single request. A style_match setting chooses between precise and flexible adherence. On fal, the same controls exist on the V4 Styles endpoints, raster and vector, and a style you created in your Recraft account works there.

OpenRouter's V4 Styles models take reference images on every request instead. On both marketplaces, the V4.1 endpoints accept colors but no style inputs, so a custom style on Recraft's newest model means calling the direct API.

Read the endpoint schema, not the model name, before you assume parity. “Recraft V4.1 on fal” and “Recraft V4.1 on the direct API” are the same model with different controls.

2. Price is the same, until it isn't

The intuitive assumption is that a marketplace charges a markup. For Recraft that is not true today: V4.1 costs $0.035 per image on the direct API, fal, and OpenRouter, and V4.1 Pro costs $0.21 on all three. Price parity at list is common because labs set the resale terms.

The differences show up at the edges. Direct API units are prepaid at $1 per 1,000 and do not expire, which suits bursty workloads. Styled calls are another example: with a saved style on the direct API or fal, you pay the $0.005 style fee once, while OpenRouter charges it on each styled request. That adds $50 per 10,000 single-image requests.

Marketplaces bill on their own terms and can reprice or retire an endpoint on their own schedule. Once you need volume pricing, a contract, or an SLA, you negotiate with the model provider either way, because the marketplace cannot sign for the lab.

3. A second company sits in your request path

Every request passes through the marketplace before it reaches the model. That adds some latency, a second party to your data-processing review, and a second status page to your incident response.

It also adds a strategic question. Beyond hosting, fal post-trains faster variants of third-party models, mostly for video, and publishes them next to the originals. Its CEO has said pretraining its own model is not on the near-term agenda.

A distributor that also ships model variants has incentives that will not always match yours or those of the labs you rely on. That is not a reason to avoid it. It is a reason to know which door you are using.

When an AI model marketplace is the right choice

Choose a marketplace when at least two of these are true:

• You are evaluating several models and want one integration to compare them.
• Your volume is low or unpredictable, and convenience is worth more than terms.
• You route between models by task and want one API shape.
• You only need base parameters: prompt, size, and a palette.
• You want one invoice across labs rather than a contract with each.

In this situation the marketplace is doing real work for you. Paying list price for aggregation is a fair trade.

When to go direct to the model provider

Go direct when any one of these is true:

• You depend on controls the marketplace schema does not expose, such as custom styles on the newest model or a managed style library.
• Your volume is high enough that terms matter more than onboarding speed. Committed-use pricing and SLAs come from the lab.
• Your compliance review needs a short data path with one named processor.
• You want a roadmap relationship: early access, feedback loops, and a person to call when a model update changes your output.
• You need the provider's full tool set. Recraft's direct API adds inpainting, outpainting, background replacement, region erase, vectorizing, and upscaling next to generation.

Going direct costs you the integration work the marketplace was absorbing. For one or two core models, that is a few days. For 10, it is a platform team.

The hybrid most production teams run

The pattern we see most often is not either-or. Teams prototype on an AI model 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 rule per model: if you call it with only a prompt, a size, and colors, leave it on the marketplace. If you need your custom style on the newest model, manage a style library, or edit outputs after generation, go direct. The controls that make the output yours are the ones most likely to be missing in the middle layer.

Recraft's models are available both ways. Marketplaces such as fal and OpenRouter are good places to check whether Recraft V4.1 fits your product's aesthetic, and V4.1 Flash at $0.007 per image keeps fast iteration cheap on fal or the direct API. The direct API is where styles on V4.1, style management, the editing tools, and volume terms live.

If you work through an agent, the Recraft MCP server sits on the direct side. Get an API key and rerun your marketplace prompts on V4.1 with a style attached.

For a broader look at what the leading image APIs offer and cost, see our guide to the best AI image generation APIs in 2026.