Speed

How models trade off intelligence against speed

How speed adds up

From prompt to finished answer

First comes the wait for the first token, then the answer streams out. The blue rectangle is that stream: its area is the output length, its height is the throughput, so its width is the streaming time. Wait plus stream spans the completion time.

Latency5.0sOutput ÷ Throughput2,000 ÷ 100 = 20.0sCompletion Time25.0s
IQ vs Completion Time
Completion time is time to last token (TTLT): time to first answer token plus output tokens ÷ throughput (TPS). Drag the input slider (1K–10K tokens) and output slider (100–10K tokens) to see how TTLT changes with the workload.
Controls:
Input10K tok
Output1K tok
IQ1:1Completion Time

How to read this chart

Completion time is time to last token (TTLT): time to first answer token plus output tokens ÷ throughput (TPS). Drag the input slider (1K–10K tokens) and output slider (100–10K tokens) to see how TTLT changes with the workload.

IQ vs Latency
Latency is time to first token (TTFT): how long a model takes to begin responding. Drag the input-length slider from 1K to 10K tokens to interpolate between AA's two measured TTFT anchors.
Controls:
Input10K tok
IQ1:1Latency

How to read this chart

Latency is time to first token (TTFT): how long a model takes to begin responding. Drag the input-length slider from 1K to 10K tokens to interpolate between AA's two measured TTFT anchors.

IQ vs Throughput
Throughput is median output tokens per second (TPS). Each model's estimated IQ is plotted against how quickly it streams its answer after generation begins.
Controls:
IQ1:1Throughput

How to read this chart

Each point is a public model. The chart compares IQ against Throughput, with color showing the model provider.