Where to Get Models

Official APIs, cloud platforms, model hubs, and aggregators differ in control and responsibility.

You've picked a model. The next question is which path to wire it in through. The same model is reachable through five channels, prices can differ by several times, and the responsibility boundaries are completely different.

Signs you'll recognize:

• A platform sells it at half the official price and you're not sure it's safe to use.
• You've wired up three providers — three SDKs, three accounts, three invoices.
• An outage hits and you don't know who to call, or where the backup is.

Five channels

Official provider API. Straight to the source. You get new models first, the clearest rate limits and SLA, and official support when something breaks. The cost: every provider needs its own account, prepayment, and settlement — and using overseas providers from China adds network and compliance friction. claude-pricing

Cloud model platforms. Alibaba Cloud Bailian and Volcano Ark are essentially model supermarkets on the cloud: one account calls many models, both first-party and third-party hosted. Their biggest advantage isn't price; it's the enterprise procurement path — contracts, invoices, compliance qualifications, and integration with the cloud services you already use. For a company going to production, this is usually the smoothest route. aliyun-bailian volcengine-ark

Aggregators. OpenRouter and SiliconFlow give you one key for dozens of models. Their value is side-by-side comparison and quick switching: evaluating five models on your task doesn't require five accounts. Good for the exploration phase and small products; before scaling up, review their stability and settlement model. openrouter

Self-hosted gateway. With an open-source gateway like One API, you run your own layer in front of official accounts you open at each provider, and expose one unified interface internally. It's the only path that gives you official prices and unified management at once: keys in one place, usage per team, one-click model switching. The cost is running that layer yourself. Once the team grows, this route has the best cost-performance. oneapi

API resellers. Third-party resale at prices well below official. Covered separately.

About resellers

This one needs explaining, because the low price is very tempting.

A reseller buys upstream quota and sells it to you. The price can sit below official because of volume discounts, exchange rates, or sources that are less than proper.

Three risks you need to know. First, your data flows through a third party you don't know. Request contents are fully visible to it — user privacy, internal documents, business data. Second, there's no stability guarantee. If the upstream account is banned, your service dies, and you have no channel to appeal. Third, this kind of resale usually violates the upstream's terms of service, so the pipeline can be cut at any time.

The conclusion: running non-sensitive experiments through a reseller is your own call. But if user data or a production service is involved, don't go down this road. The money you save doesn't compare to a single data incident.

Picking by stage

• Exploration — you're still comparing models: aggregator. One key to try everything.
• Small-scale validation — a feature just shipped, low volume: official API or cloud platform. Open one account, get the pipeline running, and learn your real costs before committing.
• Going to production — contracts, invoices, compliance: cloud platform. The bottleneck here is usually procurement and legal, not technology.
• At scale — multiple projects, multiple models, central control: self-hosted gateway on top of official accounts. Key management, usage attribution, and model switching become the real problems.

Whatever path you take, do one thing from day one: wrap model calls in your own layer. Don't let business code scatter SDKs from every provider. Switching, backups, and fallbacks then touch one place — and switching will eventually happen.