CO2 Methodology
How the History tab's CO2 chart is calculated, what its red dotted reference lines mean, and the studies behind every number. Nothing here is measured — no AI provider exposes per-request energy telemetry — so treat every figure as an informed estimate, not a fact.
The CO2e Estimate
Every session's CO2e footprint comes from capacity.EstimateCO2Grams, which multiplies a session's total tokens by a per-model coefficient. Which coefficient it uses depends on whether the model's own vendor has published one:
| Tier | Meaning | Formula |
|---|---|---|
| Provider-disclosed | A vendor or peer-reviewed lifecycle assessment already publishes a gCO2e-per-token figure, grid mix and datacenter overhead baked in. | grams = tokens × gCO2ePerToken |
| Fallback | No model-specific figure exists (true for Claude, GPT, Llama, and most others). Falls back to a cross-model energy estimate converted through a global-average grid intensity and datacenter efficiency. | grams = tokens × 0.0003 Wh × 1.15 PUE × 460 gCO2/kWh ÷ 1000 |
Coefficients & Sources
| Model family | Tier | Coefficient | Source |
|---|---|---|---|
| Gemini | Provider-disclosed | 0.00003 gCO2e/token | Google, Measuring the environmental impact of delivering AI at Google scale (Aug 2025): 0.24 Wh / 0.03 gCO2e median text prompt, normalized to ~1,000 tokens. PDF |
| Mistral / Mixtral | Provider-disclosed | 0.00285 gCO2e/token | Mistral AI, Carbone4 & ADEME lifecycle assessment (Jul 2025): 1.14 gCO2e per 400-token prompt, full lifecycle including amortized training. Announcement |
| Everything else (Claude, GPT, Llama, etc.) | Fallback | 0.0003 Wh/token | Epoch AI's revised estimate of ~0.3 Wh per frontier-model query (Feb 2025), divided by an assumed ~1,000 tokens/query — the tier most sessions land in, since neither Anthropic nor OpenAI publishes per-token energy figures. Article |
The fallback formula's other two constants: a 1.15 PUE (datacenter power-usage-effectiveness overhead), the midpoint of EcoLogits' disclosed 1.1–1.2 hyperscaler range (EcoLogits), and a 460 gCO2/kWh global-average grid carbon intensity, from Ember's 2025 Global Electricity Review (report).
This is always an estimate, never a measurement
No provider exposes real per-request energy telemetry, and hosting-path family names are inconsistent across providers (Gemini can appear as
gemini, vertex_ai-language-models, or a Bedrock alias depending on how it's served). Irrlicht matches on the model name itself and always surfaces the tier alongside the number rather than presenting it as precise.
Reference Lines: Everyday Equivalents
A raw gram count is hard to judge on its own, so the CO2 chart overlays up to 3 red dotted horizontal lines, each at the height of an everyday activity's CO2e footprint — picked to span the visible range so a session's footprint reads as "about a car mile" or "about a third of a flight" rather than an abstract number. Lines are chosen automatically from the table below and scale with the chart's zoom level; a chart showing a few grams won't draw a flight line.
Like the estimate itself, these equivalents fall into two confidence tiers: some come from a government methodology or a primary study; others are physics-based ballpark figures with no single authoritative source. Both are labeled below.
| Equivalent | Estimated CO2e | Confidence | Source |
|---|---|---|---|
| A web search | 0.2 g | Primary source | Google's own 2009 estimate. Official Google Blog — no updated official figure has been published since; efficiency gains since 2009 mean the true figure today is likely lower. |
| Charging a smartphone | ~10 g | Order-of-magnitude estimate | No authoritative single source. A ~10–20 Wh battery, plus charging losses, at the global-average grid intensity (~460 gCO2/kWh, Ember 2025) works out to roughly 2–30 g depending almost entirely on local grid mix. |
| 1 hour of video streaming | 36 g | Primary source, contested | IEA's 2020 global-average midpoint. IEA stresses this varies by roughly 100x depending on device, network, and grid — treat it as a rough midpoint, not a constant. |
| Boiling a kettle | ~60 g | Order-of-magnitude estimate | No authoritative single source. Physics baseline (~0.1–0.12 kWh to boil ~1L of water) at varying grid intensity works out to roughly 15–85 g depending on country. |
| Driving 1 km by car | 170 g | Primary source | UK Government (DESNZ), GHG Conversion Factors for Company Reporting 2025 — average petrol/diesel car, tank-to-wheel. Conversion factors |
| 1 kWh of average grid electricity | 460 g | Primary source | The same 460 gCO2/kWh global-average grid intensity used in the fallback CO2e formula above. Ember, Global Electricity Review 2025. Report |
| A hot shower | ~1 kg | Order-of-magnitude estimate | No authoritative single source. Heating ~50–80 L of water by ~25°C for an 8–10 minute shower takes roughly 1.5–2.3 kWh; at the ~460 gCO2/kWh global-average grid intensity (Ember 2025) that works out to roughly 0.6–1.1 kg depending on shower length and water-heater type. |
| A load of laundry | ~1.5 kg | Order-of-magnitude estimate | No authoritative single source. Widely cited figures range from ~0.6 kg (cold wash, line-dried) to ~3.3 kg (hot wash, tumble-dried); 1.5 kg represents a mixed-practice midpoint. |
| Burning 1 liter of petrol | ~2.35 kg | Primary source | US EPA, Greenhouse Gas Equivalencies Calculator: "8,887 grams of CO2 emissions per gallon of gasoline consumed," converted to metric (8,887 g ÷ 3.785 L/gallon ≈ 2,348 g/liter). Methodology & references |
| Manufacturing a bicycle frame | ~5.5 kg | Order-of-magnitude estimate | No single authoritative figure — steel-frame lifecycle assessments put a single frame anywhere from ~5.5 kg to ~15 kg CO2e depending on study boundaries and alloy; 5.5 kg represents the lower, frame-only end of that range. |
| Manufacturing a pair of running shoes | ~9.5 kg | Primary source | MIT Materials Systems Laboratory (Kirchain & Olivetti, 2013): manufacturing-phase emissions of 9.5 ± 2.7 kg CO2e per pair, roughly 68% of the shoe's ~14 kg total lifecycle footprint. MIT News |
| A pair of jeans, cradle to grave | ~33.4 kg | Primary source, contested | Levi Strauss & Co.'s own lifecycle assessment of a pair of 501 jeans: ~33.4 kg CO2e across the full cradle-to-grave lifecycle (fiber, manufacturing, and years of consumer washing/drying), of which ~16.2 kg is manufacturing alone. Case summary. Other independent studies report figures as low as ~12.6 kg for a narrower system boundary — treat this as a wide-range estimate, not a fixed constant. |
| A short-haul flight (London → Paris) | ~43.8 kg | Primary source | DEFRA/DESNZ 2025 short-haul, economy, radiative-forcing-inclusive factor (125.76 gCO2e/passenger-km) × ~348 km. Older studies (e.g. a 2006 Eurostar-commissioned figure of 122 kg) put this route ~3x higher, likely reflecting older, less efficient aircraft. Conversion factors |
| A tree's CO2 absorption for a year | ~60 kg | Primary source | US EPA's equivalencies-calculator methodology: an urban tree's average annual uptake over its first 10 years (mixed conifer/deciduous), citing McPherson et al. 2016. Methodology & references. A smaller, commonly cited ~22 kg/year figure comes from the Arbor Day Foundation, which doesn't publish its underlying methodology — we use EPA's documented figure instead. |
| A month of average car commuting | ~118 kg | Derived estimate | Derived, not a separate primary figure: ~15 miles/day round trip × 393 g/mile (US EPA's 2022 average fleet fuel economy, 22.8 mpg) × 20 workdays ≈ 118 kg. Methodology & references |
| A laptop, cradle to grave | ~185 kg | Primary source | Apple's own 13-inch MacBook Pro Product Environmental Report (Nov 2020): 185 kg CO2e across the device's full lifecycle, ~73% from production. Other Apple laptop models range roughly 150–325 kg CO2e depending on size and generation. Product Environmental Report |
| A long-haul flight (London → New York) | ~650 kg | Primary source | DEFRA/DESNZ 2025 long-haul, economy, radiative-forcing-inclusive factor (117.04 gCO2e/passenger-km) × ~5,555 km. Other calculators put this route anywhere from ~590 kg to ~1 tonne depending on load-factor and lifecycle assumptions. Conversion factors |
| A round-trip long-haul flight (there and back) | ~1.3 t | Derived estimate | Simply 2 × the long-haul figure above (650 kg × 2 ≈ 1.3 t) — not a separate source, included so the ladder doesn't jump straight from a single long-haul flight to multi-tonne car/person figures. |
| An average car's emissions for a year | ~4.29 t | Primary source | US EPA, Greenhouse Gas Equivalencies Calculator: "4.29 metric tons CO2e/vehicle/year" for a typical passenger vehicle. Methodology & references |
| An average person's annual carbon footprint | ~4.8 t | Primary source, high variance | Global average, not a single-country figure — deliberately chosen to stay relatable worldwide. Our World in Data: "4.8 tonnes per person." Per-capita CO2 emissions. National averages vary hugely (some countries well under 1 t/year, others well over 10 t/year). |
| Roughly 2 average cars' annual emissions | ~8.58 t | Derived estimate | A simple multiple of the average-car-per-year figure above (2 × 4.29 t = 8.58 t), not a separate source — included to keep the ladder dense between a person's and several cars' annual footprint. |
| Roughly 3 average cars' annual emissions | ~12.87 t | Derived estimate | A simple multiple of the average-car-per-year figure above (3 × 4.29 t ≈ 12.87 t), not a separate source. |
| Roughly 6 average cars' annual emissions | ~25 t | Derived estimate | A simple multiple of the average-car-per-year figure above (25 t / 4.29 t ≈ 5.8), not a separate source — included to keep the ladder dense near the top of the range. |
| Roughly 21 people's average annual carbon footprint | ~100 t | Derived estimate | A simple multiple of the average-person-per-year figure above (100 t / 4.8 t ≈ 20.8) — a generous ceiling for the reference-line ladder; real sessions, even aggregated org-wide over a year, are extremely unlikely to approach this scale. |
A lower-carbon comparison
For contrast: the same 2025 UK conversion factors put a national rail trip at roughly 44 g CO2e per passenger-km (including well-to-tank) and a coach/long-distance bus at roughly 34 g — both well below the ~170–215 g/km for an average car, and far below flying.
Caveats & Uncertainty
- Every figure above is a widely-cited public average, not a measurement of irrlicht's own AI provider traffic or your own local habits.
- Flight, tree, and car figures assume specific routes, load factors, or growth conditions spelled out in the table — a real trip or a real tree will vary.
- The smartphone, kettle, shower, laundry, and bicycle-frame figures have no single authoritative source; they're physics-based ballpark ranges, not government statistics.
- The jeans figure varies significantly across LCA studies (~12.6–33.4 kg) depending on system boundary; we use Levi's own full cradle-to-grave number rather than a narrower manufacturing-only figure.
- Grid carbon intensity varies enormously by country and even by time of day — anywhere from near-zero (hydro/nuclear-heavy grids) to well over 700 gCO2/kWh (coal-heavy grids), against the ~460 g global average used in the fallback formula.