We are not trying to be your router.
LiteLLM, Portkey, Kong, and OpenRouter are data planes: they forward, retry, and observe, and the good ones do it well. The Gateway is an inference control plane: it decides each request, prices it, proves it, and learns your organization. You still get the table stakes: one endpoint, retries and failover to another provider or model, telemetry, user and group management, and budgets. The reason to come here is what no data plane attempts.
- data plane
- forwards, retries, observes
- control plane
- decides, prices, proves, learns
- reroutes
- only after you accept
- cost
- reconciled to your invoice
- fee
- a share of proven savings
What no data plane attempts
- Routing you can audit, moved only by you. Our recommendations carry parity evidence graded on your own traffic, and nothing reroutes until you accept. A data-plane router moves traffic on price or latency, silently.
- Savings checked against the invoice. Data planes recompute cost from a price table they maintain. The Gateway reconciles spend and savings to the provider invoice you actually pay, and our fee is a printed slice of savings we proved there.
- Your organization's knowledge, enforced at the door. Approved sources ground every answer with citations, and a question outside them gets a refusal, not a guess. Your own procedures, runbooks, escalation steps, support policy, ride into every request that needs them. Elsewhere, grounding lives in app-level RAG tools and procedures in copy-pasted prompts.
- Screening that routes on proof, not promises. Regulated data is screened deterministically on every channel, and a data class routes only to a provider whose recorded attestations cover it: health data without a signed BAA on file has nowhere to go, and the refusal is recorded.
A data plane makes traffic move faster. A control plane decides whether it should move, what it should cost, and what proof should exist afterward. We compete on the decision, not the pipe.
The receipts, vendor by vendor
Table stakes we share, and they often do well:
- One endpoint over many providers, retries, and failover: every product on the list. So do we.
- Telemetry, usage reporting, and spend dashboards: Portkey, Helicone, and OpenRouter are strong here. You get robust telemetry from us too, and every report also reconciles to the invoice.
- Org hierarchy, budgets, and key management: LiteLLM and OpenRouter handle these well. So do we, with self-service keys inside admin-set ceilings.
- Kong and LiteLLM restore redacted values after the model replies, as we do. Reversible redaction is not unique to us, and we never claim it is.
- NotDiamond routes on quality measured from your own eval data. As routers go, it is the closest to our evidence bar, without the acceptance step.
- Credal and Glean ground answers in org sources with per-user permissions, and do it well, as enterprise AI platforms you build inside. No transparent gateway grounds every client behind one URL, which is the seat this product takes.
Where the category difference shows:
- OpenRouter's BYOK plan is free up to $25,000 a month of list-price usage, then charges 5 percent of the equivalent OpenRouter cost. That is a fee on spend, not on savings.
- Portkey gates its PII anonymizer and BAA behind enterprise plans. Here, every safeguard ships on every plan. Plans differ in evidence and cadence, never in safety.
- Cloudflare AI Gateway estimates cost and accepts per-request overrides. Kong and Portkey compute from price tables an operator maintains. None reconciles to the invoice you pay.
- LiteLLM documents that budget enforcement fails open when its database is down. Our budgets act before the request leaves: warn, throttle, downgrade, refuse.
Screening, honestly:
- On breadth of detection classes, LiteLLM with Presidio is wide, and that is real. Ours differs in where it reads: messages, system instructions, tool calls, query parameters, and documents with extractable text. And in being deterministic: the same input under the same policy version finds the same identifiers on replay.
- The Gateway blocks a route on a missing attestation for the data class, health data without a signed BAA, for example, and records the refusal.
When a data plane is the right buy
- All traffic in one cloud, policy met by its gateway. Keep the cloud gateway. Come back when you add a second provider.
- You want a free, self-hosted proxy and you run it well. LiteLLM is a fine data plane. Keep it, and add the Gateway underneath when you need the decisions, the proof, or the reconciliation.
- Your job is offline evaluation and prompt tuning. Use an eval platform. If you adopt the Gateway later, feed that platform its capture export.
- You want the cheapest route per call and nothing else. A runtime router does that. It will not ask first, and it will not prove what it saved.
Check us the same way.
We admit customers in cohorts, in waitlist order. Don't miss the next one. When your slot comes up, you start on one real workload, on your plan from day one. Every claim on this page is testable in that run, against your own invoice. If we are not earning our fee, leaving is the same URL change, reversed.