AI API benchmarks

AI API response times, measured daily

Time to first token from real servers around the world

AI APIs
5
Test locations
6
Latest data
10 Oct 2026

What time to first token means →How we measure →

Current AI latency benchmarks

Typical and tail latency for popular AI APIs, measured from 6 global locations.

  • ChatGPT (Low reasoning effort)

    gpt-5.6-sol

    Typical TTFT990 ms
    p95 TTFT1,788 ms
    Geographic spread1.8×
    14-day trendDaily typical time, 27 Sept to 10 Oct: between 775 ms and 1,180 ms.
    Fastest locationMontreal
  • Claude (Low reasoning effort)

    claude-opus-5

    Typical TTFT1,132 ms
    p95 TTFT1,794 ms
    Geographic spread1.2×
    14-day trendDaily typical time, 27 Sept to 10 Oct: between 999 ms and 1,431 ms.
    Fastest locationMontreal
  • DeepSeek (Reasoning turned off)

    deepseek-v4-flash

    Typical TTFT661 ms
    p95 TTFT882 ms
    Geographic spread1.4×
    14-day trendDaily typical time, 27 Sept to 10 Oct: between 558 ms and 711 ms.
    Fastest locationSingapore
  • Grok (Low reasoning effort)

    grok-4.6

    Typical TTFT2,649 ms
    p95 TTFT4,878 ms
    Geographic spread1.2×
    14-day trendDaily typical time, 27 Sept to 10 Oct: between 2,335 ms and 3,095 ms.
    Fastest locationSan Francisco
  • Mistral (Provider default settings)

    mistral-medium-latest

    Typical TTFT330 ms
    p95 TTFT580 ms
    Geographic spread2.0×
    14-day trendDaily typical time, 27 Sept to 10 Oct: between 302 ms and 375 ms.
    Fastest locationAmsterdam

*Low reasoning effort: ChatGPT, Claude, Grok.

**Provider default settings: Mistral.

***Reasoning turned off: DeepSeek.

Listed alphabetically. APIs with different marks were sent different reasoning settings, so their times aren't a like-for-like comparison.

Fast from where?

Darker cells are slower than that API's fastest location.

Typical time to first token in milliseconds, by provider and test location
AmsterdamSan FranciscoMontrealSingaporeTokyoMumbai
ChatGPT1,0259287489771,0031,326
Claude1,1051,0731,0431,2051,1581,251
DeepSeek700728734527622589
Grok3,0262,4742,5912,6562,7442,641
Mistral214333277435435326
Typical TTFT in ms. Compared with each API's fastest location:

Where the wait comes from

Reaching the server is at most 6% of the wait. The rest is the model.

These bars come from one test of 20 requests per city on 27 Sept, not from the two weeks of daily requests behind the typical times, so they won't match those exactly.

Latest findings

How we measure

Every day, each API gets the same one-line prompt as a streaming request from every test location. We time how long the first visible word of the reply takes. Hidden reasoning doesn't count, and we don't measure how fast the rest of the reply arrives or how good it is.

Read the full methodology →

How fast is your own API?

Run a free test from the same 6 cities and see where the time goes.

Test your API

Free, no account needed.