Every gateway can fan out to multiple models. Only one also compresses your input tokens, caps spend per key before a leak becomes an invoice, and writes the saving back on every call. Twenty-three capabilities, graded honestly — including where the others are ahead.
Graded on out-of-the-box capability as commonly deployed. Partial means the capability exists but needs configuration, a plugin, or covers only part of the ground.
| Capability | NeuroRoute | RouteLLM | OpenRouter | LiteLLM | Portkey |
|---|---|---|---|---|---|
| Routing intelligence | |||||
| Multi-provider from one API | Yes | Partial | Yes | Yes | Yes |
| Automatic cost-based routing | Yes | Yes | Partial | Partial | Partial |
| Task-aware, quality-bar routing | Yes | Partial | — | — | Partial |
| Learns from your own feedback | Yes | Partial | — | — | — |
| Auto-escalates on empty/failed responses | Yes | — | Partial | Partial | Partial |
| Self-hosted models (vLLM) | Yes | Yes | — | Yes | Partial |
| Cost governance | |||||
| Keep your own keys (BYOK) | Yes | Yes | Partial | Yes | Yes |
| Per-key + org $ spend quotas | Yes | — | Partial | Yes | Yes |
| Input-token compression | Yes | — | — | — | — |
| Semantic + prompt caching | Yes | — | Partial | Partial | Yes |
| Per-request savings proof | Yes | — | Partial | Partial | Partial |
| Transparent tier pricing, no % of savings | Yes | Partial | Partial | Partial | Partial |
| Data trust & control | |||||
| PII detection & routing policy | Yes | — | — | Partial | Partial |
| Input & output guardrails | Yes | — | — | Partial | Yes |
| Request-log search + OpenTelemetry | Yes | — | Partial | Yes | Yes |
| SSO, workspaces & scoped keys | Yes | — | Partial | Partial | Yes |
| Multi-workspace membership & key assignment | Yes | N/A | — | Partial | Partial |
| Per-member least-privilege scoping | Yes | N/A | — | Partial | Partial |
| Person + workspace-level audit trail | Yes | N/A | — | Partial | Partial |
| Provable zero-retention (per response) | Yes | Partial | Partial | Partial | Partial |
| Encryption at rest + CMEK | Yes | N/A | — | Partial | Partial |
| Crypto-shred erasure + GDPR export | Yes | — | — | — | Partial |
| MCP access (OAuth 2.1) | Yes | — | — | — | — |
RouteLLM (LMSYS) is an open-source strong/weak routing model; OpenRouter is a hosted multi-model aggregator; LiteLLM is an open-source multi-provider proxy with budgets & caching; Portkey is a commercial AI gateway with guardrails. All are capable tools — this reflects out-of-the-box capability as commonly deployed and will vary by version and configuration. Corrections welcome.
Three rows carry most of the gap. They are the ones that show up on the invoice and in the security review, not in a benchmark.
No other tool in the matrix ships it. Up to ~40% fewer input tokens before the request ever reaches a provider — savings that compound on top of routing.
Multi-workspace membership, least-privilege scoping and a person-level audit trail. Most gateways stop at an API key.
A flat tier fee and a published per-million rate. The incentive to route you cheaply is not diluted by taking a cut of the result.
