What was measured
The i6eal AI Latency Index records time to first content token (TTFT) from a Lambda function in Frankfurt. It compares the medians of daily measurements across two completed UTC weeks: 12 to 18 October 2026 and 19 to 25 October 2026. Only weekdays with a value in both weeks count.
The observed changes
For Mistral Small (route: gateway), the TTFT median rose from 344 to 443 milliseconds. For GPT-5 mini (route: direct, OpenAI), it fell from 824 to 685 milliseconds. Both comparisons rest on seven paired days each.
Importantly, the changes apply only within each individual endpoint. You should not derive a provider ranking or a general statement about a vendor's performance from them. The index deliberately does not rank direct and gateway routes against each other.
Interpretation with caution
The measurement is a short synthetic probe, not a measurement of your entire workload. The collector discards a warm-up run and keeps the faster of two successful probes per measurement round. The observed differences establish no cause, statistical significance, model quality or reliability.
For your own planning this means: if response time matters to you, measure with your real prompts on your own use case before basing decisions on these two values.
Observed latency changes
| Endpoint | Route | Paired days per week | Before TTFT (ms) | Current TTFT (ms) | Change |
|---|---|---|---|---|---|
| Mistral Small (Mistral) | gateway | 7 | 344.00 | 443.00 | +28.78 % |
| GPT-5 mini (OpenAI) | direct | 7 | 824.00 | 685.00 | -16.87 % |
2026-10-12–2026-10-18 → 2026-10-19–2026-10-25 (UTC).
Coverage and measurement limits
7 endpoints have sufficient paired days and unchanged model/route descriptions. 0 endpoints have insufficient paired days; 0 were excluded because model/route descriptions changed.
TTFT measures time to the first content token from a Lambda in Frankfurt. This compares medians of daily measurements in two completed UTC weeks. Only weekdays with a value in both weeks count; each endpoint needs at least five paired days. The collector discards a warm-up and keeps the faster of two successful probes per measurement round. This is a short synthetic probe, not a measurement of your entire workload. Direct and gateway routes are not ranked against each other. Observed differences establish no cause, statistical significance, model quality or reliability.
i6eal KI-Latenz-Index. Source timestamp: 2026-10-26T01:20:59.141Z.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
