On LiteLLM’s own AIGatewayBench, Cortega added 0.33 ms p99 per request — with policy, data protection, and MCP tool checks running on every call. LiteLLM’s published figure for its Rust gateway is 0.66 ms.
agentgateway (Cortega’s data plane) benchmarked against LiteLLM on identical hardware: Fortio, mock backend, 32 connections, June and August 2026.
| Gateway | Requests / sec | p99 latency added | Memory |
|---|---|---|---|
| Cortega data plane (agentgateway) | ~36,000 | ~2 ms | ~25 MB |
| LiteLLM, Rust mode | 983 | 71 ms | 2.15 GB |
| LiteLLM, Python mode | 3,198 | 32 ms | 11.8 GB |
A separate machine sends traffic over a real network. The upstream model is held at a fixed delay; the request rate rises until the gateway runs out of room.
| Requests per second | Success | Median latency | Added by the gateway |
|---|---|---|---|
| 40,000 | 100% | 62 ms | ~2 ms |
| 48,000 | 100% | 63 ms | ~3 ms |
| 52,000 | 100% | 65 ms | ~5 ms |
| 53,000 | 99.97% | 69 ms | out of headroom |
Cortega runs as a fleet, so capacity grows as you add gateways. Full method on the performance page.
None of this is on LiteLLM's roadmap. It ships as part of the same platform, not a bolt-on.
Industry and agentic-security benchmark suites — jailbreaks, prompt injection, multi-turn escalation, MCP tool poisoning — scored against your live guardrails.
See AI Bench →Social engineering, financial fraud and risk scoring, and policy-document compliance, alongside PII/PCI and jailbreak detection.
See Guardrails →Ranked upgrade candidates from your own measured traffic, and a way to benchmark one before it ever serves a live request.
See Model Manager →Tenant telemetry is isolated on read today; the architecture is designed to run a large tenant's analytics on its own store as volume grows.
See MSP governance →Cortega Agents scores live agent sessions for risk, compliance, quality, and cost — and Endpoint Guard governs the coding assistants your developers run.
See Coding Assistant Governance →LiteLLM is a strong developer gateway. Some of what an enterprise needs sits behind its paid Enterprise tier; some it does not do at all.
| Capability | LiteLLM | Cortega |
|---|---|---|
| Model access and routing | Open source. One API across providers, keys, budgets, routing, fallbacks. | Included, with policy applied on every route. |
| SSO and RBAC | Paid. Free up to 5 users, then an Enterprise license. | Included. Tied to policy, audit, model access, and tool actions. |
| Audit logs | Paid. Listed as an Enterprise feature. | Included. Links identity, policy result, data, model, and outcome. |
| Guardrails | Open source. Set per request, key, or team. | A full AI firewall: block, redact, route, approve, or record before the call proceeds — plus social engineering, fraud/risk, and policy-document guardrails. |
| MCP tool governance | Partial. MCP support, credential work on the roadmap. | Full: servers, tools, permissions, calls, evidence, and policy in one model. |
| Model recommendations & migration testing | Not offered. | Ranks upgrade candidates from your measured traffic, then benchmarks a candidate before it serves production. |
| Security & industry benchmarks | Not offered. | AI Bench: jailbreak, prompt-injection, MCP tool-poisoning, and industry suites, scored against your live guardrails. |
| Coding-assistant governance | Not offered. | Endpoint Guard configures Claude Code, Codex CLI, OpenCode, and Pi on enrolled machines with one governed key, one switch to disable. |
| Employee / shadow AI | Not offered. Governs application and developer API traffic. | Endpoint Guard covers ChatGPT, Claude, browsers, and internal assistants. |
| Where AI traffic data sits | In its database. Requests and spend logged to Postgres. | Kept out of the operational database. Governance does not store prompts or responses. |
| Usage intelligence | Limited. Spend and usage tracking. | What teams use AI for, where quality drifts, where strategy and real use differ. |
| Multi-tenant analytics at scale | Basic multi-tenant support via virtual keys. | Tenant-isolated policy and gateways today; architected to shard analytics per tenant as volume grows. |
Cortega’s 0.33 ms figure is our run of AIGatewayBench, LiteLLM’s own gateway-overhead benchmark, one request stream at a time with the load generator on the gateway node. LiteLLM’s published AIGatewayBench figures put its Rust gateway at 0.66 ms, Portkey at 2.29 ms, and Bifrost at 4.54 ms of added latency per request (p99).
The throughput, latency, and memory comparison is from the agentgateway project’s published benchmarks (agentgateway.dev, 26 June 2026 and 13 August 2026): Fortio, mock backend, 32 connections, agentgateway 1.4.0 versus LiteLLM 1.98.0 in Rust mode.
Cortega’s 52,000 req/s figure is a separate test over a real network against a fixed-delay upstream, on a single 8 vCPU instance, using the Bifrost benchmark harness.
Foundation is free — one gateway, standard guardrails, no time limit.
Tell us what you’re evaluating and we’ll run Cortega against it on your terms.