Time to first token, usually shortened to TTFT, is the wait between pressing send and the moment a model produces its first token. It shapes your first impression of how fast an AI product feels: some tools sit silent for three or four seconds before anything appears, while others start streaming almost instantly. The difference usually lives in this opening window, not in how fast the model writes once it gets going.
What the model is doing before the first word
Many people assume a model computes its answer word by word from the very start. In reality there is a concentrated processing step first: the system reads your question together with the whole conversation and any attachments, works out how every part relates to the rest, and waits for compute capacity. Only after that does the first token appear. This step scales with input length — the longer the material you paste in and the thicker the chat history, the longer the model spends "reading the question", even if your final question itself is one short sentence.
TTFT and streaming speed are two different bills
Once the first token is out, the model switches to generating one token after another, and what matters then is tokens per second. That produces some interesting combinations: some models start instantly but slow down as they write, so long answers drag; others pause for a few seconds and then flood the screen, finishing sooner overall. Judging a product's speed only by how quickly text starts appearing often leads to the wrong verdict — for short questions TTFT dominates, while for long documents and reports, total completion time matters more.
What stretches TTFT noticeably
First, very long context: uploading dozens of pages or pasting a whole codebase means the model spends real time just reading. Second, reasoning models: they think internally before speaking, and the more thinking, the longer the wait — which is why deep-thinking modes generally start slowly. Third, peak-hour queues and cold starts: if the same product feels sluggish in the evening and snappy at dawn, the model has not changed; there are simply more requests ahead of yours.
Two common misconceptions
One: a fast first token means a strong model. TTFT mostly reflects scheduling and input processing, not answer quality; smaller models often start faster but answer more shallowly. Two: a long wait means it has frozen. If the interface shows the model is thinking, or you just uploaded a large file, waiting a bit longer is usually normal processing; only a long, unchanging stall that survives a refresh points to a connection or service problem.
How to judge speed in practice
When comparing products or troubleshooting slowness, split the timing in two: send-to-first-token is TTFT, and first-token-to-finished is generation time. If the first part is too long, shorten the context, drop unnecessary attachments, or avoid peak hours; if the second part is too long, look for a model that streams faster. Splitting it this way makes it clear which step is actually slow.
