ModelPilot vs. AI gateways & routers
If you're comparing us to AI gateways (OpenRouter incl. Fusion, Portkey, Helicone, LiteLLM, Cloudflare AI Gateway) or LLM routers (Martian, Not Diamond, Unify, Requesty, RouteLLM) — here's the honest version. We're not a gateway, and we optimize differently from the routers too.
The three things gateways structurally don't do
- Your prompts never leave your system. A gateway sits in your data path — it has to see your prompt to route and meter it. ModelPilot classifies locally; only a task category and numeric features ever reach us. Never prompt text, model outputs, or your API key.
- You pay only for realized savings. Gateways bill on usage (token markup, a credit fee, or a % of spend / BYOK fee) — they make money when your spend goes up. We charge a share of the savings we actually deliver. No savings, no bill — our incentive is your bill going down.
- Proven cost reduction, not just access. We measure savings against a held-out control arm and check the cheaper model side-by-side for non-inferiority on your own traffic (0% false-downgrades on our golden set). The number on your dashboard is audited, not a marketing figure.
Side by side
| Typical AI gateway | ModelPilot | |
|---|---|---|
| What it is | Hosted API in front of many providers/models | Drop-in proxy on your own Claude account |
| Goal | Access & breadth (and, for OpenRouter Fusion, higher quality via a multi-model panel — ~4–5× the cost of one call) | Cut cost — route down to the cheapest good-enough model |
| Your prompts | Flow through their servers (routed, sometimes sampled/logged) | Never leave your system — classified locally |
| Billing | Token pass-through + fee, or % of spend / BYOK fee | % of realized savings (20% PAYG / 15% on subscription tiers) |
| Relationship | They become your API + billing layer | You keep your direct Anthropic account; we fail open to it |
| Proof | Quality/latency optimization | RCT control arm + non-inferiority judging, per-category floors |
| Scope | Broad — many providers | Deep on the Claude family |
What about LLM routers (Martian, Not Diamond, RouteLLM)?
Routers also pick a cheaper/better model per request — and several are excellent. But to decide, a router has to read your prompt: it routes on the prompt, so your content is in its data path. ModelPilot routes on a local classification instead — the decision is made on your box from a task category + numeric features, so the prompt itself never reaches us. We also bill on realized savings (routers bill usage or enterprise contracts) and prove it with a held-out control arm and non-inferiority checks on your own traffic, rather than quoting a savings range.
When a gateway or router is the better fit
We'd rather be straight with you: if you need one API across many providers (OpenAI, Google, open-weight models), built-in fallbacks across vendors, or — with OpenRouter Fusion — a multi-model ensemble for maximum answer quality, a gateway is the right tool, and we're not trying to replace it. Some teams even run both: a gateway for breadth, ModelPilot to stop overspending on the easy requests.
ModelPilot is the better fit when you're Claude-heavy, your prompts are sensitive (they can't leave your environment), and you want cost cut with proof — paying only when it works.
Head-to-head
Detailed comparisons: ModelPilot vs OpenRouter · ModelPilot vs Martian.
Questions? krethikram@gmail.com · How we optimize without seeing your data