Direct API vs Cloud Marketplace vs Aggregator: LLM 2026
The Short Answer
There are three ways to buy model access in 2026, and the choice matters more than most teams realise:
- Direct API — buy from the lab. Cheapest list price, day-one access to new models, no intermediary. One vendor relationship per model family.
- Cloud marketplace (Bedrock, Vertex, Azure/Foundry) — buy inside your existing cloud account. Governance, data residency and unified billing, at roughly a 10% premium and a lag on new releases.
- Aggregator (OpenRouter and similar) — one API key across many vendors. Fastest model switching, built-in failover, an extra dependency in the path.
Most mature stacks end up with two of the three, not one.
Head-to-Head
| Direct API | Cloud marketplace | Aggregator | |
|---|---|---|---|
| Price vs list | List | ~+10% regional premium | List + small margin |
| New model access | Day one | Days to weeks later | Usually day one |
| Model breadth | One vendor | Curated catalogue | Widest |
| Switching cost | High (new contract, new SDK) | Low within catalogue | Lowest |
| Governance / audit | Vendor-specific | Strongest | Weakest |
| Data residency | Vendor regions | Your cloud regions | Depends on route |
| Procurement | New vendor each time | Existing cloud contract | One vendor |
| Failover | Build it yourself | Within catalogue | Built in |
| Extra dependency | None | Cloud provider | Aggregator |
Last verified: August 2026.
Price: The Comparison Is Not What It Looks Like
Direct is cheapest on paper. Cloud marketplaces carry a regional premium of roughly 10% over direct rates, and vendor-specific modifiers stack on top — Anthropic’s US inference geo, for example, applies a 1.1× multiplier across all token categories.
But list price is not spend. Two effects routinely invert the ranking:
Aggregators arbitrage. When a model is served by several providers, an aggregator can route to the cheapest healthy one per request. Over a month that can beat any single direct relationship, margin included.
Marketplaces come with committed-spend discounts. If you already have a large cloud commitment, model spend may draw down against it at negotiated rates that direct list pricing cannot match. For big enterprises this frequently makes the “expensive” option the cheapest one.
The honest conclusion: run the numbers on your actual volume. Anyone quoting a universal winner on price has not.
Risk: Each Route Fails Differently
This is the dimension that 2026 made vivid.
Direct API risk is concentration. If the vendor reprices, retires the model, or restricts access, you have no in-path alternative. You also carry a separate contract, separate billing and separate security review per vendor — which is why teams under-diversify.
Cloud marketplace risk is lag and catalogue control. New models arrive later, sometimes much later, and you get whatever your cloud has negotiated. Your model roadmap becomes a function of your cloud provider’s partnership decisions.
Aggregator risk is the added hop. You have one uptime dependency, one billing relationship and one company’s business continuity between you and every model you use. Aggregators are also themselves acquisition targets, which — as 2026 demonstrated across the tooling layer — means their commercial terms can change under new ownership.
The pattern worth internalising: negotiated contracts routinely contain change-of-control clauses. Any intermediary in your path is a company whose acquisition can alter your access. That is not hypothetical any more.
Where Each Route Wins
Direct API wins when you need a model on release day, you are large enough to negotiate directly, or you need capabilities that only surface on the vendor’s own endpoint — new tool-calling features, beta parameters and experimental variants usually land direct first and propagate outward.
Cloud marketplace wins when governance is the binding constraint. Data residency in your own cloud regions, audit trails your compliance team already accepts, IAM integration, and a procurement path that does not require a new vendor review per model. In regulated industries this beats a 10% premium every time.
Aggregator wins when model diversity and switching speed are the point. Prototyping across ten models, routing by cost tier, or running a cheap-model-first architecture with frontier escalation — all far easier behind one key. It is also the cheapest way to keep a warm fallback, since your second-choice model is a string change rather than a new integration.
The Combination Most Teams Land On
Almost nobody stays on exactly one route once they are past prototype.
Startups: aggregator as primary, one direct relationship with whichever vendor dominates their spend. The aggregator gives breadth; the direct relationship gives day-one access and better rates at volume.
Mid-size: direct primary for the main model, aggregator as the warm fallback. This keeps list pricing on the bulk of spend while making failover a config change.
Enterprise: cloud marketplace primary for governance, direct secondary for models the marketplace lags on. Aggregators often fail procurement review here, and that is a defensible call.
The anti-pattern is single-route purism. A stack that can only reach models one way inherits every failure mode of that way, with no mitigation.
How to Decide
- Name your binding constraint — price, governance, or model breadth. Only one of them is actually binding; the other two are preferences.
- Price your real volume on all three routes, including commitment discounts you already have.
- Ask the thirty-day question: if this route stopped working in thirty days, what would you do? Write the answer down.
- Wire the second route before you need it. Integrated and idle beats theoretical and urgent.
- Re-check quarterly. Marketplace catalogues, aggregator margins and direct pricing all move faster than annual planning cycles.