The fastest LLMs: output speed and time to first answer, measured
Which AI models stream fastest and answer soonest, measured on live traffic, and why reasoning effort turns a two-second model into a two-minute one.
Speed has two parts and most leaderboards show only one. Output speed is how fast tokens stream once the answer starts, in tokens per second. Time to first answer is how long you wait before the answer starts at all, and for reasoning models that includes the whole thinking phase. A model can stream at two hundred tokens per second and still make you wait a minute. Numbers as of 9 Sept 2026.
Fast and good
Raw speed is easy: small models are fast. The useful list is models above a quality floor. These stream at a hundred tokens per second or more with an index above 55:
| # | Model | Output speed | Index | Price $/M |
|---|---|---|---|---|
| 1 |  Gemini 3.7 Flash medium Google | 282 tok/s | 64.4 | $1.50 |
| 2 | Gemini 3.8 Flash medium Google | 273 tok/s | 63.5 | $1.50 |
| 3 |  Muse Spark 1.2 Meta | 219 tok/s | 62.6 | $2.00 |
| 4 | Gemini 3.5 Flash medium Google | 209 tok/s | 63.9 | $3.38 |
| 5 | Gemini 3 Flash thinking Google | 200 tok/s | 61.5 | $1.13 |
As of 9 Sept 2026. Full list.
Quick to answer
For interactive use, the number that matters is how long until the first word of the answer. These models start within two seconds:
| # | Model | Time to first answer | Index | Price $/M |
|---|---|---|---|---|
| 1 |  Qwen3.5 Plus thinking Alibaba | 1.1 s | 57.2 | $1.29 |
| 2 |  Claude Sonnet 4.5 high Anthropic | 1.4 s | 57.8 | $6.00 |
| 3 |  Kimi K2.5 Moonshot AI | 1.6 s | 56.1 | $1.20 |
| 4 |  DeepSeek V3.2 thinking DeepSeek | 1.8 s | 55.9 | $0.315 |
As of 9 Sept 2026. Full list.
Where the measurements come from
Output speed comes from Artificial Analysis, which runs every model through the same harness and reports the median. Time to first token comes from OpenRouter's live traffic, the median over the last thirty minutes at the moment of our daily fetch. Time to first answer, the figure that includes reasoning, comes from Artificial Analysis. We keep those families separate on purpose: mixing a first-token measurement with a first-answer one produces nonsense like a model that "answers" in four seconds and also in four minutes.
Reasoning effort is the whole story
Look at any frontier model's configuration table. At low effort it answers in a few seconds; at max effort the same model takes minutes before the first token of the answer, because it is thinking first. The scores go up with effort and so does the wait. If you are building something interactive, the max configuration that tops the quality board is probably not the one you want to deploy, and the compare page will show you exactly what you trade away by stepping down.
Provider choice matters too
The same open-weights model can stream at three times the speed on one host as on another. Every model page has a "where to run it" table with each provider's measured throughput, first-token latency and price, so you can pick the host, not just the model.
What speed does to cost
The per-task boards also estimate time: first answer plus output tokens divided by speed. A summary of a 30-page report takes seconds on a fast model and minutes on a slow reasoning one. For batch work that is a throughput bill; for a person waiting, it is the product.