Simulate real-time token streams to compare AI response pacing and perceived responsiveness. It is a browser player: no login, and no model API call.
About 4 characters count as 1 token, so you can feel the pace. That is not the model’s own tokenizer. Rows are grouped by family: OpenAI, Anthropic, Google, SpaceXAI Grok, DeepSeek, Qwen, GLM, Kimi, MiniMax, and speed specialists. Cost is shown only when an output price is published. Gemini 4 Argon has no published median, so it is not listed.
On the 2026-10-12 snapshot, the fastest is Celeris-1 (1,532 t/s), then Mercury 2.5 (711.8 t/s) and Gemini 3.5 Flash-Lite (419.6 t/s).
How to use it
- Pick a model line, such as GPT or Claude, then a version. Play uses that median tokens/s.
- Press Play and watch the text come out at that pace. You can edit the text or paste a public page link.
- Turn on Compare and set another row as B to race them. Switching family clears that pane.
Questions
What is Token Speed Simulator?
Simulate real-time token streams to compare AI response pacing and perceived responsiveness. See the pace of AI.
Does playback send the text to a model?
No. Playback types in the browser at the speed you picked. It does not log you in or call a model API. A public link is read for its text; if the browser cannot read it, a reader proxy is tried.
Where do the speeds come from?
They are Artificial Analysis median output tokens/s, snapshotted on 2026-10-12. They are for feel, not a live test of your account.
Which model is fastest right now?
On the 2026-10-12 snapshot, the fastest is Celeris-1 (1,532 t/s), then Mercury 2.5 (711.8 t/s) and Gemini 3.5 Flash-Lite (419.6 t/s).
How do I compare two models?
Turn on Compare, pick A, set another row as B, and press Play once. The same family keeps the text when only the speed changes. A different family clears that pane.
How is cost calculated?
When AA publishes an output price, spend is about tokens times that price, and the rate is about tokens/s times that price. Missing prices stay a dash. Nothing is invented.
Which languages are available?
Simplified Chinese, Traditional Chinese, English, Japanese, Korean, Spanish, Portuguese, Romanian, French, and German. Auto follows the browser. A specific language is saved in the bubufu-language cookie and opens that language URL.
Languages
Source: Artificial Analysis · 2026-10-12. Page updated 2026-10-13.