TTFT / TPS

You might say

Why do some AI tools take forever to show the first word while others answer right away?

Measures of time to first output and of token generation speed after output beginsTTFT measures time from sending a request to receiving the first token. TPS measures tokens produced each second afterward. A quick first token does not guarantee a short total response time.

Know first

Streaming Response Token

TTFT & TPS

Where do three kinds of wait come from on one timeline?

These times and speeds are illustrative; they only compare how the metrics relate.

Send requestReceive first TokenAnswer complete

TTFT 1.2sInitial wait

40 TPSTokens generated each second after it starts

Total time 6.2s

B starts speaking sooner, but generates more slowly afterward, so its full answer finishes later.

Separate first output, speed, and total time

Wait for TTFT: TTFT runs from send to the first text token.

Then watch TPS: TPS is tokens per second after generation begins, not transactions per second.

Then measure total time: A longer answer can still take longer to finish even when it begins quickly.

Quick checkChoose the best answer

Model A shows its first token in one second but generates slowly; Model B starts in three seconds and finishes quickly. How should speed be compared?

Measure time to first content, generation rate after start, and total completion separatelyCompare only which model shows the first character and call it faster overallCombine network waiting and generation into one TPS number without phases

You can say this to an AI Agent

Measure TTFT, TPS after output begins, and total completion separately under the same model settings and answer target. Do not replace full comparison with feels instant, and record network, queue, and generation phases.

Learn next

1Streaming Response→2Token→3Token Usage & Cost→

Want GPT or Codex?VibeHubGPT support

GPT-6 Astra

Build in 3D. Put your computer to work.