Provider Claims vs. Reality
We ran the same benchmark on every provider. Here's how their published numbers compare to what we measured.
Methodology: TTFT measured via streaming from Ottawa, Canada and US East. 5 runs per model. Test prompt: "Explain the concept of neural networks in one paragraph." STT: 5-second WAV test audio. TTS: 92-character sentence. Measurements from April 2026. Results vary by region and load.
LLM Inference — TTFT P50
| Provider | Model | Their Claim | Our TTFT P50 | Source | Status |
|---|
Speech-to-Text — End-to-End Latency P50
| Provider | Model | Their Claim | Our Latency P50 | Source | Status |
|---|
Text-to-Speech — TTFB P50
| Provider | Model | Their Claim | Our TTFB P50 | Source | Status |
|---|
Found an error? [email protected]