Tokens per Second Calculator
Work out how long a model takes to write a response from the output length, generation speed in tokens per second and time to first token, with the words-per-minute equivalent.
Last updated: October 2026
Total response time
20.5 seconds
Formula below · 5 sources (nngroup.com, doi.org, help.openai.com, platform.claude.com, ai.google.dev) · Updated Oct 2026
Compare with similar
About this calculator
A model’s response time has two parts: the wait before the first token appears (time to first token, TTFT, which grows with prompt length, queueing and any hidden reasoning) and the time to write the rest at its generation speed:
response time = TTFT + output tokens ÷ tokens per second
Tokens per second depends on the model, the provider’s hardware and load, and for local models on your GPU and quantization; measure it from a few real requests (most APIs and local runners report it). Reasoning tokens count as output, so a model that thinks for 2,000 tokens before a 500-token answer generates 2,500 tokens.
To make the speed concrete, tokens are converted to words with each provider’s English rule of thumb, and compared with an average adult silent reading speed of 238 words per minute for non-fiction (Brysbaert, 2019). Response-time bands follow Nielsen Norman Group’s limits: about 1 second keeps the user’s flow of thought, about 10 seconds their attention. Streaming the output makes the perceived wait closer to the time to first token.
How to use
Enter the expected output length in tokens, the model’s generation speed and its time to first token, and pick the tokenizer for the word figures. Example (the defaults): 1,000 tokens at 50 tokens/s with 0.5 s to the first token: 0.5 + 1,000 ÷ 50 = 20.5 seconds. With OpenAI’s ¾ word per token that is 750 words written at 50 × 0.75 × 60 = 2,250 words per minute, about 9.5× reading speed; reading it takes 750 ÷ 238 = 3.2 minutes. Example 2: a 300-token answer at 150 tokens/s with 0.3 s TTFT: 0.3 + 2 = 2.3 s. Example 3: a local model at 12 tokens/s writing 800 tokens after 1.5 s: 1.5 + 66.7 = 68.2 s.
Frequently asked questions
What is a good tokens per second speed?
Anything faster than people read feels fluent when streamed: at OpenAI’s ¾ word per token, about 5–6 tokens per second matches a 238 words-per-minute reader. Hosted models typically stream well above that; local models depend on the GPU, the model size and the quantization.
What is time to first token?
The delay between sending the request and receiving the first output token. It covers network time, queueing, reading the prompt (longer prompts take longer) and, for reasoning models, thinking before the visible answer starts.
How do I measure tokens per second?
Divide the output tokens reported in the response’s usage by the time from the first to the last streamed token. Local runners such as llama.cpp and Ollama print the generation speed after each response.
Related calculators
Sources & references
- Nielsen Norman Group: Response Times, the 3 Important Limits
- Brysbaert (2019): How many words do we read per minute? Journal of Memory and Language
- OpenAI Help: What are tokens and how to count them
- Anthropic: Claude models overview (context windows, words per token)
- Google: Understand and count tokens (Gemini API)
Think a result is wrong? Report a problem with this calculator.