Cost monitoring without a proxy — an honest comparison
Every approach has tradeoffs. Here is when a proxy is the right choice, and when it is not.
When a proxy is the right tool
- Per-request logging. You need to see every prompt and completion, attribute them to users, or tag them with metadata. A proxy captures this at the request level because it sits in the request path.
- Prompt inspection. You want to scan inputs for PII, enforce content policies, or log prompts for compliance. A proxy can read and transform the payload before it reaches the provider.
- Real-time rate limiting. You need to reject requests immediately when a budget is exhausted — not after a polling interval. A proxy can enforce this per-request because it controls the path.
When a proxy is the wrong tool
- Latency matters. Every proxy hop adds latency — typically 20–100ms per request, more under load. For latency-sensitive workloads, this is a direct cost to user experience.
- Prompt exposure is a concern. A proxy sees every prompt and completion by definition. If your threat model excludes third-party prompt access, a proxy is a non-starter. Even with encryption in transit, the proxy operator can read your data at the application layer.
- You already have production keys deployed. Migrating traffic through a proxy means changing every API endpoint in your application. For teams with keys deployed across services, this is an integration project, not a configuration change.
The ModelBrake tradeoff
ModelBrake never sits in the request path. You give us a read-only admin credential — the one that can see spend but cannot call models. We poll the provider's usage API on a schedule: every 15 minutes for OpenAI, Anthropic, and Vercel, every 6 hours for Bedrock (constrained by Cost Explorer data availability).
The tradeoff is a detection gap. Spend that occurs between polls is not visible until the next poll completes. For most teams, 15 minutes is fast enough to catch a four-figure spike before it becomes a five-figure invoice. For teams that need per-request enforcement, a proxy is the right tool and ModelBrake is not.
Side-by-side
| Gateway / proxy | ModelBrake | |
|---|---|---|
| Setup | Route all traffic through proxy | One read-only admin key |
| Latency | Added hop on every call | None |
| Prompt access | Proxy sees every prompt | Never sees prompts |
| Alert speed | Real-time per request | 15-minute polls |
| Credential risk | Production key lives in proxy | Read-only key, separate from app |
Founding member rate, locked for life. Limited spots.
- OpenAI, Anthropic, Bedrock, Vercel — polled every 15 min
- Anomaly alerts + per-key spend ceilings
- Email + webhook delivery (Slack-compatible)
- Every feature we ship next, at this rate
$29/mo, billed monthly. Cancel anytime, rate stays locked.