One-pager · Studio operating decision

Build vs. buy

Every AI studio faces this stack decision three ways at once. The answer isn't one of them — it's knowing which layer each kind of work belongs on, and keeping the freedom to move as the market shifts monthly.

The call

Buy

Suites for exploration

A few seats so creatives touch every new model the week it ships. R&D spend, not production infrastructure.

Build

APIs for production

Gemini direct plus one aggregator. Every repeatable pipeline lives here. This demo runs exactly this stack.

Neither, forever

Re-evaluate monthly

Model IDs and prices in one config file. When a leader changes, it's an edit — not a rebuild.

The stack

01Creative suitesHiggsfield · Freepik · Krea$15–150 / seat / moExploration

02API aggregatorsfal.ai · ReplicateUsage-pricedProduction

03Direct model APIsGemini · Black Forest Labs · RunwayUsage-priced, negotiableAnchor

Each layer sits closer to the model than the one above it — and gives up convenience for control as you descend.

Head to head

Criterion	Creative suites	API aggregators	Direct APIs
Pricing model	Per seat + credits	Per generation✦ edge	Per generation, negotiable at scale
Access to new models	Day one, across vendors✦ edge	Days to weeks	Day one, that vendor only
Workflow integration	None — UI only	Full API✦ edge	Full API
Rights & indemnity	Vendor terms, varies	Passed through from the model owner	First-party, negotiable✦ edge
Cost at high volume	Breaks down	Linear and predictable	Linear, improves with commitment✦ edge
Switching cost	Low — cancel seats	Low — change one endpoint string✦ edge	Medium — per-vendor integration
Where it breaks	Volume scales	You need a first-party SLA✦ edge	The leaderboard flips

“Edge” marks the layer that wins that row outright — no layer wins enough of them to win the page, which is the point.

The layers in detail

Creative suites

Higgsfield · Freepik · Krea

Subscription UIs wrapping 15+ third-party models with presets and editing tools. They own no frontier models — they are a convenience layer over the same APIs anyone can call.

Wins when

Creatives explore and iterate hands-on, daily
Day-one access to every new model
Small team, modest volume

Fails when

Volume scales — per-seat plus credits beats you
You need workflow integration, QA gates, audit trails
Rights and data-governance terms need enterprise review

API aggregators

fal.ai · Replicate

One key, ~1,000 curated production endpoints (fal) or 50k+ community models (Replicate). The pragmatic middle: model diversity without one integration per vendor.

Wins when

Repeatable pipelines — versioning, adaptation, batch
You swap models often as leaders change
Cost maps linearly to output

Fails when

You need a first-party enterprise contract and SLA
A model exists only on its vendor's own API

Direct model APIs

Gemini · Black Forest Labs · Runway

First-party access: newest capabilities first, best rate limits, enterprise terms negotiated with the model owner. Gemini is the standout for retail — reasoning, vision, stills and video behind one key.

Wins when

The workload anchors on one vendor's ecosystem
Legal needs first-party rights and indemnity
Scale justifies negotiated pricing

Fails when

You over-commit and the leaderboard flips (see Sora 2's API sunset)
Every extra vendor is another integration to maintain

Own the workflow.
Rent the models.

Workflows, prompt systems, quality gates and brand knowledge compound. Models are commodities that change quarterly. Build the first, rent the second, and never let the second dictate the first.

Three questions that settle it

Who indemnifies the output?: Rights and data-governance answers only exist at the suite-contract or first-party-API level. If a brand ships it, someone has to stand behind it.
What does this cost at 500 assets a week?: Per-seat suites are flat until volume turns credits into the real bill. APIs are linear from day one — and model routing (draft cheap, finish premium) is where the savings actually live.
Could we leave in a week?: The market leader has changed roughly quarterly since 2024. Any layer you can't exit in a week is a liability, not a partnership.

What this demo actually runs on

Gemini direct for reasoning, vision, Nano Banana stills and Veo video; fal.ai for Flux, Kling, Seedance, Runway and ElevenLabs Music. Every model ID and price lives in one config file with environment overrides — the switching cost this page argues for, made real.

See the routing table →