Same prompt, different appetite.
Type anything and see how much each model would likely write back. Nothing leaves your browser and no model is called.
In plain words
Type a question and see roughly how much each chatbot would write before answering — the more it writes, the more electricity it uses.
It’s an estimate from how these models behaved on similar test questions.
Spotted: language name, code term, debugging term.
You typed ≈ 25 tokens. The answers would run 968 to 10,665 — 11× apart.
Each model's measured average on OckBench's 60 coding problems.
Each phone is one full charge (≈ 15 Wh), filled to the energy the answer takes. Tap a model for cost, carbon, water and accuracy. Add or remove models below.
Models4 of 9tap to open
The footprint is the answer, not the question. What you type is a rounding error next to what comes back, so the model you pick and its effort setting decide the energy. See the full comparison →
How this estimate is made4 notestap to open
- No model runs. Your prompt is sorted into math, coding, science or everyday by matching keywords in your browser, and each model's measured average for that type of problem is shown. Averages come from OckBench live leaderboard — per-domain results (math / coding / science): 100 math, 60 coding and 40 science problems. A real answer to your exact prompt could be shorter or longer. The gap between models is the reliable part.
- Everyday prompts (emails, chat, summaries) have no measurement. They show each model's average across all 200 hard problems, an upper bound. How much the gap narrows on easy prompts has not been measured, so we don't guess.
- Input tokens are approximated at ~4 characters per token, so no tokenizer has to load. They are left out of cost and energy: reading a prompt is far cheaper per token than writing an answer, and the answer is hundreds to thousands of times longer. Cost is output tokens × list price (BenchLM.ai — LLM API pricing comparison (Aug 2026)). Energy, carbon and water use the same bands as the token efficiency page.
- For DeepSeek V4 Flash (high), Flash (max) and Qwen3.5 9B, OckBench's per-domain file averages about 3%, 2% and 5% higher than its overall file. Both are shown as published.