WattMatters - Router Efficiency Test
Which model is right for your task?
Model access
Check models sends one tiny generation request to gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, and the separate judge gpt-4o. Small API charges may apply. The credential stays on the server. One check runs at a time.
Enter a prompt or select a Preset ICP:
Tradeoff criteria
Estimated CO₂e emissions (g per reference query)
CO₂e means carbon dioxide equivalent. This grid converts the supplied energy ranges. It is not a live-run measurement and not the separate US-grid estimate (110–640 g/kWh) shown on a comparison.
| Model | Role | Short 100 input / 300 output tokens | Medium 1,000 input / 1,000 output tokens | Long 10,000 input / 1,500 output tokens |
|---|---|---|---|---|
| GPT-4.1 | Candidate | 0.20–0.41 g | 0.93–1.29 g | 1.46–1.92 g |
| GPT-4.1-mini | Candidate | 0.13–0.19 g | 0.47–0.62 g | 0.62–0.86 g |
| GPT-4.1-nano | Candidate | 0.06–0.09 g | 0.16–0.24 g | 0.26–0.32 g |
| GPT-4o | Judge, not a candidate | 0.12–0.18 g | 0.34–0.51 g | 0.86–1.16 g |
Formula: grams CO₂e = Wh × 0.35, which is 350 g CO₂e/kWh. Display uses decimal half-up rounding to two places. The published watt-hour ranges already include the paper’s PUE 1.12, so this conversion does not multiply by PUE again.
Source: Jegham et al., Table 4, version 6. Each interval converts the supplied mean ± one standard deviation on the reference Azure infrastructure (PUE 1.12). These are estimates, not guaranteed bounds and not direct provider measurements. Values are per reference query, not per 1,000 output tokens.
GPT-4o is the separate judge in this reference catalog. Its values are not added to candidate emissions and are not an A/B/C assignment. Names in a run stay hidden until Reveal.
Reference estimates using 350 g CO₂e/kWh. Inference electricity only; excludes training and hardware manufacturing.
Energy method
Primary energy is a published reference estimate from Jegham et al., Table 4, shown as Wh per reference query. Each interval is the paper’s mean ± one standard deviation. These are not measured live electricity, not guaranteed bounds, and not energy reported by the OpenAI API.
The paper’s reference deployment is Microsoft Azure with PUE 1.12. Published workloads are Short 100 input / 300 output tokens, Medium 1,000 input / 1,000 output tokens, and Long 10,000 input / 1,500 output tokens. A live prompt can differ from those workloads. Compare uses the Medium scenario as an approximation, not a measurement of this request.
Carbon on a live comparison is a separate US-grid estimate: lower Wh × 110 / 1,000 and upper Wh × 640 / 1,000, in grams CO2e. It is not the paper’s Azure carbon estimate and not the fixed 350 g CO₂e/kWh tradeoff-criteria grid. Equivalent-query figures multiply both energy endpoints by 1,000. They are projections, not a batch benchmark.
Distinct ranges can identify a hidden model. Anonymity is limited. The catalog below is not the A/B/C assignment for a run.
Inference electricity only; excludes training and hardware manufacturing.