What is a \"good\" tokens-per-second (tok/s) speed for chatting vs. coding?
For casual chatting, 10 to 15 tokens per second (tok/s) feels comfortable—it tracks right around normal human reading speed. For coding and AI agents, aim for at least 30 to 50+ tok/s. When writing software, you are usually waiting on entire blocks of code to finish rather than reading along word by word.
In AI, 1 token equals roughly ¾ of an English word (100 tokens ≈ 75 words). Code burns through tokens much faster because syntax symbols, brackets, and indentation each count separately. Because you consume prose and code differently, the right hardware target depends heavily on your daily workflow:
| Speed Tier | Real-World Feel | Best For |
|---|---|---|
| 2 – 5 tok/s | Slow trickle (CPU-only territory). Noticeable wait for anything longer than a few sentences. | Overnight batch jobs, quick 1-sentence sanity checks, or background summaries. |
| 10 – 15 tok/s | Human reading pace. Words stream in just fast enough to read comfortably in real time. | Everyday Q&A, brainstorming, email drafts, and casual chat. |
| 30 – 50 tok/s | Snappy and responsive. Whole paragraphs and functions appear within a couple of seconds. | Interactive coding, refactoring functions, and heavy daily writing. |
| 80 – 150+ tok/s | Near-instantaneous. Answers flash onto the screen almost immediately. | Inline autocomplete, reasoning ("thinking") models, and autonomous coding agents. |
Why coding demands 3x to 5x more speed than chatting
- You don't read code word-by-word: In a chat, you start reading the moment the first sentence appears. In an editor, you sit idle until the model finishes a 60-line function so you can copy, paste, and run it. At 10
tok/s, a 500-token code block leaves you staring at your screen for nearly a full minute; at 50tok/s, it wraps up in 10 seconds. - Agents and "thinking" models burn hidden tokens: Reasoning models (like DeepSeek R1 or Qwen thinking modes) and agentic tools (like Cline, Roo Code, or OpenCode) generate thousands of hidden "thought" tokens and tool calls behind the scenes before printing a single line of final output. A sluggish 10
tok/ssetup turns one agentic task into a five-minute coffee break. - Prompt processing (Time to First Token) matters just as much: Before generating a reply, your system has to ingest your prompt and any attached codebase files. A dedicated GPU chews through prompts at 500 to 2,000+
tok/sfor an instant start, whereas CPU-offloaded rigs often stall for 15 to 30 seconds before the very first word appears.
Keep in mind when comparing hardware benchmarks that published numbers are usually measured on short prompts. Once your conversation stretches past 15,000+ words of context, real-world generation speed naturally drops by 20% to 40%.
Looking for verified hardware rigs?
Check our real-time £/GB VRAM index to find systems optimized for local model inference.