susty: sustainability advice that knows what it costs

Got questions on what's more sustainable?

Ask about diet, travel, home energy, or anything else.

How we counted this

Estimates, not promises. Provider token counts are measured where available; energy, device power, location, and some grid values are modelled, inferred, or derived.

This conversation

Carbon

This chat so far	0.00 g CO₂e
Low: our estimate. High: if AI uses as much energy as the upper end of published estimates (Oviedo et al. 2026), 2.7 times ours. A plausible spread, not a margin of error.
Answers	0
Average per answer	—
Cut off or failed (in the total, not counted as answers)	0
Your device's grid	471 g/kWh, life-cycle
Servers' grid (assumption: US data centres)	307 g/kWh, basis not stated by source

Energy

Answered by	—
Tokens sent (context)	0
Tokens written back	0
of which thinking	0 tokens
AI inference, after PUE ×1.12	0 Wh
chips reading, 0.05 mWh/token	0 Wh
chips writing, 0.5 mWh/token	0 Wh
Hosting & network	0 Wh
Your device	0 Wh
Electricity, this chat	0 Wh
Share of a phone charge	0%

Water

Cooling at the data center	0.00 mL
At power stations	0.00 mL
Water, this chat	0.00 mL
Plus hydro reservoirs, if all charged to power	0.00 mL

Water used up (evaporated), not counted in the carbon. Reservoirs are kept out of the total because they also supply farms and towns.

Your grid (for your device)

Answer with

Reading this on

The latest answer on other models

Appears after the first answer.

Your history on this device

Today	0.00 g CO₂e, 0 answers
Last 7 days	0.00 g CO₂e, 0 answers
All saved	0.00 g CO₂e, 0 answers

Saved only in this browser; nothing is sent anywhere. Chats with at least one question are kept until you clear them, oldest dropped first if space runs out. Each entry keeps the figures it was priced with, under the methodology version it was priced with; nothing is repriced. Central figures; the export has the high end too.

What to do about it

Offsetting this costs under a hundredth of a cent.

At about $320 per tonne, the 2024 average price of durable carbon removal, this conversation is a rounding error. Each of these three actions covers it many times over:

Wash clothes in cold water, around 200 g saved per load.
Thermostat down 2°F, roughly 300 kg a year in a typical heated home.
One beef meal a week swapped, on the order of 200 kg a year.

To put it in perspective

The footprint of asking is small next to the footprint of what you do with the answer.

Notes

We do our best to make these calculations as accurate as possible. Some of it has to be estimated, so here's where the numbers come from and where we're guessing. The full method has all of it.

One thing we don't know precisely, but would be great to find out: which specific servers are answering your questions, in a data center somewhere else, running on their own electricity. Your device is counted on your own grid, above. For the servers we assume a typical US data center: weighted by capacity, US data centers sit on grids averaging about 307 g CO₂e per kWh, according to Watershed (September 2026). Watershed doesn't say whether that figure counts only what power stations emit or the full life cycle, so we say so too. We'd love to know exactly which servers are doing the work and what that data center draws, but that isn't public information right now. Our goal is to run this locally on a sustainable server ourselves.

The large meter is the carbon of this chat so far, and it goes up with every answer. The ledger also shows the average per answer: that total divided by the answers that finished normally. Failed or cut-off attempts stay in the total when their energy can be measured or modelled, but they do not count as answers. The total includes AI inference, hosting/network, and the whole device you're reading on during estimated attended time. Typing and scrolling are evidence of attention; after a period without activity we stop counting device time. Longer chats also cost more per answer because the conversation context is sent back each time.

Methodology

Functional unit: one completed exchange (an "answer" on screen) One user prompt plus the corresponding response that reaches normal provider completion and is committed to the conversation. It is reported in the ledger as a session average: total operational session CO₂e divided by completed exchanges. The headline is the running session total.
Energy per token (modelled assumption) Susty currently models 0.05 mWh per input token and 0.5 mWh per output token. These coefficients are assumptions, not direct measurements from the cited literature, and they sit at the low end of published estimates: provider figures and research for a comparable question run from about the same to several times higher. That is why the ledger shows this chat as a range: the high end assumes AI energy at the upper quartile of one published estimate (Oviedo et al., 0.60 Wh for a standard question), about 2.7 times ours. Model differences are scaled from EcoLogits' estimates of each model's size, using a simplified form that likely understates how much more large models use. Luccioni, Jernite & Strubell (FAccT 2024) is background on AI inference energy: Power Hungry Processing: Watts Driving the Cost of AI Deployment?
Data centre overhead PUE ×1.12 (modelled assumption) Susty uses 1.12 as a hyperscale-data-centre assumption. It is not the Uptime Institute global average; the serving facility is unknown.
The servers' grid: about 307 g/kWh (assumption) AI inference and hosting are priced on a typical US data center's grid, not yours: Watershed's capacity-weighted average for US data centers, about 307 g CO₂e/kWh (September 2026). Which data center serves a request is not public, and Watershed does not state the accounting basis of that figure. The grid sources below apply to your device only.
Great Britain: live, by region NESO Carbon Intensity API, half-hourly, so GB readings here are current rather than averaged. The national figure is NESO's published actual; for the 14 distribution regions NESO publishes only a forecast for the current half hour, which is what we use. Open and keyless; nothing about you is sent with the request.
United States: hourly by region EIA-930 hourly generation by fuel type, turned into an intensity with EIA's own per-fuel figures (coal 2,257 lb/MWh, gas 976 lb/MWh). Two caveats worth knowing: EIA publishes hourly CO₂ only in bulk spreadsheets, so this is derived from the generation mix rather than read off, and the feed itself runs about a day behind. We therefore take the same hour on the previous day, which tracks the daily curve far better than their newest figure would. That one is the middle of last night. If no matching hour is published yet, we use the most recent complete hour and say so. Interregional imports are ignored. Needs a free EIA key; without one we fall back to the annual figures below.
United States: annual, by eGRID subregion EPA eGRID2023 Rev 2 CO₂e output rates, converted from lb/MWh, are used for subregion fallbacks. Subregions vary widely, which is why Susty prefers a narrower region where possible.
Everywhere else: national annual averages Ember's 2024 national figures. Annual means, so they miss both time-of-day swings and regional spread. Ember counts life-cycle emissions (including fuel supply chains), while the GB and US sources above count only what power stations emit, so these run somewhat higher on the same grid.
Carbon removal priced at about $320/tonne CDR.fyi: the weighted-average price of durable carbon removal sold in 2024, down from about $490 in 2023. The often-quoted $100 per tonne is the US Department of Energy's long-run target, not today's price.
One smartphone charge is about 19 Wh US EPA's figure (0.019 kWh), from US Department of Energy data, including what the charger loses, so it is more than the battery itself holds. The burger comparison is a quarter-pound patty at Poore & Nemecek's (2018) global average for beef from dairy herds; beef from beef herds would be about three times higher. The tea and shower figures are order-of-magnitude estimates we have not traced to a primary source.
Dark mode display power (3–9% at normal brightness) Dash & Hu, MobiSys 2021: How much battery does dark mode save? An accurate OLED display power profiler for modern smartphones. Measured across several OLED phones: 39–47% of total phone power at 100% brightness, but only 3–9% at the 30–50% brightness in normal use.
Whole-device power and attended time (modelled / estimated) The browser can't measure your device's actual power draw or verify human attention. Susty models whole-device power from the selected device and panel, then accrues it during estimated attended time based on page visibility and recent interaction. The OLED phone theme delta is calibrated into the 3–9% band above; laptop and monitor theme effects are extrapolated.