The same backend, built two ways. From scratch, an agent writes a server, wires routes, rolls its own role-based auth, installs dependencies, debugs, and deploys. On a Magic Cloudlet it's described in one sentence and comes out live, hosted, and RBAC-secured at the runtime. Drag the assumptions to size the gap for your team.
The benchmark is a real build: a tasks and clients database joined by a foreign key, a full CRUD API of eight endpoints, and access locked to a single role. From scratch that is roughly 140,000 tokens of generation and debugging. On a cloudlet it was about 25,000 — an ~80% reduction, because the two genuinely expensive parts, authentication and hosting, collapse into the prompt and the runtime.
| Path | Input tokens | Output tokens | Cost |
|---|---|---|---|
| From scratch | |||
| Magic Cloudlet | |||
| Saved |
Savings % = the share of work a cloudlet can streamline × the efficiency gain on it — about 45% × 80% ≈ 36% of total token spend by default, and ~80% on the streamline-able slice itself. Cache-hit and rate changes move the dollar figures but not that percentage.
Token counts are calibrated estimates, not metered telemetry. Only the backend and integration slice of a developer's day gets the cloudlet discount; general coding, writing, and analysis are excluded. Engineering-time savings are not counted here and typically exceed the token savings. Rate basis: premium frontier model metered pricing — $10/M input, $50/M output, $1/M cached input.
20 minutes. We'll walk the validation boundary against your use case and tell you plainly whether a cloudlet fits — and what it would save.