The marketplace for every model you actually ship.

What "marketplace" actually means in 2026

The word marketplace is overloaded. Hugging Face is technically the largest, with 45,000+ model cards uploaded by anyone — research labs, forks, fine-tunes, community experiments. The signal-to-noise ratio is great if you are a researcher and bad if you are shipping a SaaS. Replicate is a model marketplace in a different sense: it gives you a per-second GPU runtime to host any model someone has packaged in Cog. AI Generate is a marketplace in the third sense: a curated catalog of 307 production-grade models, priced per unit, behind a single Bearer token, with a playground UI for testing and a webhook surface for async generation. Every model in our catalog has a named upstream provider (OpenAI, Anthropic, Google, Black Forest Labs, ByteDance, Suno, ElevenLabs, fal, Replicate-hosted) and a per-unit price you can quote your own customers from.

Why one credit pool changes the build vs buy maths

When your stack reaches the third or fourth provider, the operational tax of running them in parallel becomes the dominant cost. You have separate invoices to reconcile, separate spend caps to monitor, separate dashboards to log into, separate Stripe webhooks to verify if you bill customers downstream. AI Generate folds all of that into one credit pool with one API key. Your finance team gets one line item; your engineers stop branching error-handling on which provider died this morning; your operators see the same dashboard whether the call routed to Claude, Gemini or Veo. The 10–40% markup pays for the consolidation, and the auto-tier discount makes the unit-price math competitive once your spend crosses ~$200 a month.

Browse, test, ship — what a model page actually does

Every model in the marketplace has a dedicated page at /model/<slug>. The page shows the per-unit price (in credits and USD), the average latency, the supported parameters, a sample request, and a one-click "open in playground" button. The playground at /playground runs the same call you will run from your backend — same model, same prompt format, same output. You can copy the resulting curl request straight into your code, or switch the SDK tab between Python and JavaScript. Onboarding becomes: log in, claim $10, open a model page, hit playground, copy curl, ship. The interactive surface is why developers pick a marketplace over wiring three to five provider SDKs into their own backend before they have shipped anything.

How curation changes when a new SOTA model lands

When a new flagship model ships — Veo 3.2, Claude Opus 4.7, GPT-5.6, the next Suno — we benchmark it against the current category leader on latency, output quality and unit price, and add it to the catalog within days. Old or replaced models stay listed for backward compatibility (so existing integrations do not break) but are marked "superseded" on the model page. The curation policy is explicit: every active model in the catalog is one we would recommend to a customer building today. There is no SEO-padding with abandoned forks or weak rebrands. Coverage breadth is a side effect of provider quality, not a vanity metric.