Flow AI vs the alternatives
There are three ways to route inference. They all give you an endpoint — the difference is what they optimize for. Flow optimizes for the cheapest model that completes the task, billed at true cost.
Direct API
Call one provider's model yourself. Full control, but you pay list price, pick the model by hand, and have no failover when it's down or slow.
Token routers
OpenRouter-style: send a chosen model across its hosts and pay a markup on tokens. Cheaper sourcing, but it never knows whether your agent's task actually finished.
Flow AI
Managed, completion-aware routing: the cheapest model that completes the task, billed at true pass-through cost + 2.5% — plus a community capacity network you can earn from.
Side by side
| Flow AI | Token routers | Direct API | |
|---|---|---|---|
| Pricing model | Pass-through cost + 2.5% | Markup on token prices | List price (no routing) |
| Routes by | Whether the task completes | Token price / availability | N/A — you pick |
| Routing | Cheapest model that completes, per request | Pick a model, route across its hosts | You pick + call one model |
| Quality cascade (escalate only if needed) | Yes | No | No |
| Prompt-cache-aware billing | Yes — billed at the cache rate | Varies | Yes (per provider) |
| Failover across providers | Yes — automatic | Yes | No |
| Community capacity & earn (Hive) | Yes — share spare capacity, earn credits | No | No |
| Drop-in for OpenAI / Anthropic / Codex | All three | OpenAI-style | Native only |
“Token routers” covers OpenRouter-style services that route a chosen model by token price and availability.
Completion > cheap tokens
A model that narrates a plan instead of calling the tool is worse than useless — and a token-router will happily send your agent there because it's cheap per token. Flow measures whether the work actually got done and routes to the cheapest model that finishes the job, escalating only when it has to. You get the savings of a router with the reliability of picking the right model.