---
title: "Benchmarks"
url: "/benchmarks/"
kind: "hub"
summary: "Compare Arion throughput and latency across HTTP forwarding, MCP tool calls and custom request processing with Wasm filters and external processors."
---

> Arion site index: [/llms.txt](/llms.txt)

# Benchmarks

Compare Arion throughput and latency across HTTP forwarding, MCP tool calls and custom request processing with Wasm filters and external processors.

Arion brings nginx-level forwarding performance to a fully programmable gateway. These benchmarks compare HTTP forwarding, MCP tool calls and request processing under load.

Read about the [test hardware and methodology](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/testbed.md) behind these measurements.

## HTTP forwarding

Explore HTTP/1.1, TLS and HTTP/2 forwarding, connection reuse and concurrency. The comparison includes Arion, nginx, Envoy, agentgateway and HAProxy.

Read the analysis: [HTTP baselines and caveats](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/baseline.md) · [Four-CPU deployments](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/baseline-small-scale.md).

## MCP gateway

Compare Arion and agentgateway across REST-backed tool mappings and upstream MCP servers, with throughput and tool-call latency.

Read the analysis: [MCP forwarding](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/mcp-forwarding.md) · [REST-backed tool mappings](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/mcp-gateway-input-schema.md) · [Native MCP upstreams](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/mcp-gateway-mcp-upstream.md).

## Dynamically program the request path with custom code

Run custom Wasm filters or call external processors. Compare throughput and latency with custom code executing on every request.

## Run the benchmarks

Run the [benchmark harness](https://github.com/arion-gateway/arion-bench) on your own hardware, or explore the [results and analysis](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/README.md#analysis) from this campaign.

Published full-scale measurements used 24 proxy CPUs; the reproduction configs now use 32. See the [reproduction notes](https://github.com/arion-gateway/arion-bench-results/blob/main/202608_bench/analysis/testbed.md#reproducibility) for details.