bubufuUnderstand output pacing with a reproducible 200-token example, then distinguish a browser simulation from a real API benchmark.
Fu Jam · Updated 2026-10-19
Separate the two kinds of waiting
Time to first token (TTFT) measures the wait before output begins. Tokens/s describes the subsequent output rate. Higher throughput does not necessarily mean an earlier start. Queueing, network conditions, prompt length and reasoning can all affect the experience.
A calculation you can reproduce
Assume an output of exactly 200 tokens at a constant rate. At 20 tokens/s it takes about 10 seconds; at 50 it takes 4 seconds; at 100 it takes 2 seconds. Divide 200 by the rate. This excludes TTFT and is an arithmetic example, not a model benchmark.
Compare pacing in the simulator
Open the player, enter identical text, enable side-by-side comparison and play at different speeds. Check the text after switching model families, as a pane may be cleared. Notice reading comfort as well as completion time. The player estimates roughly four characters per token rather than using each model’s tokenizer.
When you need an actual API test
For purchasing or deployment, use your own representative tasks. Fix the model version, parameters, input and output limit, repeat the test and record TTFT, streaming duration, actual usage and failures separately. Use provider-reported token usage for billing, not the player’s estimate.
Know where text goes
Playback of typed text does not call a model. Importing public URLs reads the target page and may use a reader proxy. Never import URLs containing login tokens. An actual API benchmark sends prompts to your chosen provider; review its data handling before using confidential text.