The fastest LLMs, ranked

Speed has two components that behave very differently. Time to first token is the pause before anything appears, and it is what users experience as lag. Tokens per second is how fast text streams once it starts. A model with excellent throughput and a slow cold start still feels sluggish in a chat interface, while a model that starts instantly can feel fast even at modest streaming rates.

Reasoning models complicate the picture further. Hidden thinking tokens are generated before the visible answer begins, so a model can post a high tokens-per-second figure while taking far longer than a slower-streaming competitor to actually deliver a response. If your product is interactive, measure wall-clock time to a complete useful answer, not the raw rate.

These figures are measurements against live endpoints, which means they move with provider, region, hardware, quantisation and time of day. The same model served by two providers can differ substantially, and a heavily loaded endpoint at peak hours will not match a benchmark run at 3am. Treat the ordering as reliable and the absolute numbers as indicative.

For interactive products, speed deserves as much weight as quality — an answer that arrives after the user has given up has failed regardless of how good it was. For batch pipelines, throughput barely matters compared with concurrency limits and cost per completed task, and it is usually a mistake to trade accuracy for it.

The pragmatic pattern is tiering: a fast model for the first response or the interactive path, and a slower frontier model behind it for the work that genuinely needs the extra reasoning. Most products that feel quick and smart at the same time are doing exactly this.

More on how these numbers are produced in the methodology and what each evaluation measures in the benchmark guide. Spotted a score that disagrees with its source? Tell us.