ngrok vs VercelAI Gateway Benchmark Comparison
ngrok was faster in 3 of 4 shared AI gateway benchmarks, and 1 was roughly even.
- Shared benchmarks
- 4
- Last run
- October 30, 2026
Scorecard
Median of each benchmark's primary metric; lower is better. Medians within 5% count as roughly even, and a benchmark gets no speed verdict when either provider completed fewer than 50% of runs.
| Benchmark | ngrok | Vercel | Result | Ratio |
|---|---|---|---|---|
| Anthropic Cold end-to-end latency · Run October 30, 2026 | 452ms 100% success | 495ms 100% success | ngrok faster | 1.10x |
| OpenAI Cold end-to-end latency · Run October 30, 2026 | 753ms 100% success | 735ms 100% success | Roughly even | 1.02x |
| Gemini Cold end-to-end latency · Run October 30, 2026 | 1852ms 100% success | 1966ms 100% success | ngrok faster | 1.06x |
| Kimi Cold end-to-end latency · Run October 30, 2026 | 1862ms 100% success | 2209ms 100% success | ngrok faster | 1.19x |
Anthropic
Run October 30, 2026Time for a cold request routed through the gateway to Claude Haiku 4.5 to return its complete response.
Cold end-to-end latency
ngrok had a 1.10x lower median cold end-to-end latency (452ms vs 495ms for Vercel).
Median cold end-to-end latency over time
11 runs from August 21, 2026 to October 30, 2026.
OpenAI
Run October 30, 2026Time for a cold request routed through the gateway to GPT-5.4 Mini to return its complete response.
Cold end-to-end latency
The medians are within 5% of each other (753ms vs 735ms), so this counts as roughly even.
Median cold end-to-end latency over time
11 runs from August 21, 2026 to October 30, 2026.
Gemini
Run October 30, 2026Time for a cold request routed through the gateway to Gemini 3.6 Flash to return its complete response.
Cold end-to-end latency
ngrok had a 1.06x lower median cold end-to-end latency (1852ms vs 1966ms for Vercel).
Median cold end-to-end latency over time
11 runs from August 21, 2026 to October 30, 2026.
Kimi
Run October 30, 2026Time for a cold request routed through the gateway to Kimi K3 to return its complete response.
Cold end-to-end latency
ngrok had a 1.19x lower median cold end-to-end latency (1862ms vs 2209ms for Vercel).
Median cold end-to-end latency over time
10 runs from August 28, 2026 to October 30, 2026.
Methodology
- One workload, every provider. The same script, payload, and model hit every provider, addressed the way its own API expects. Nobody gets a tuned or special path.
- Repeated on a schedule. Every suite runs many iterations per provider, automated on a recurring schedule, so results reflect steady state rather than one lucky request.
- Raw results published. Every run is committed as JSON to the public benchmarks repo, next to the code that produced it. Nothing is cherry-picked or held back.
- How this comparison is built. Each benchmark compares the two providers' medians; results within 5% count as roughly even. A comparison is published only when both providers completed the latency benchmark on at least two of the same model families reliably: at least 50% success for each provider, in benchmark runs from the last 30 days.
Frequently asked questions
Have another question? Email us.
ngrok was faster in 3 of 4 shared AI gateway benchmarks, and 1 was roughly even. In Anthropic (October 30, 2026), ngrok had a median cold end-to-end latency of 452ms versus 495ms for Vercel, 1.10x faster. In OpenAI (October 30, 2026), ngrok had a median cold end-to-end latency of 753ms and Vercel 735ms, which counts as roughly even. In Gemini (October 30, 2026), ngrok had a median cold end-to-end latency of 1852ms versus 1966ms for Vercel, 1.06x faster. In Kimi (October 30, 2026), ngrok had a median cold end-to-end latency of 1862ms versus 2209ms for Vercel, 1.19x faster.