Tracker+ puts powerful Anthropic models to work behind Discover, headteacher and CEO reports, parents’ evening briefs and the assessment engine. As adoption accelerates, we measure exactly what that AI costs the planet — and we’re deliberate about matching the right model to each task. Every piece of analysis a school receives carries roughly the carbon of a single mug of tea.
If you genuinely care about something, you start — humbly — by measuring it. So before making any claim about responsible AI, we counted: every token, every model, every month. The right question isn’t whether AI uses energy — everything does — but how much useful work you get for each gram of carbon, and whether you are honest about it. On that measure Tracker+ is deliberately, measurably efficient. A large part comes down to one discipline: we run powerful Anthropic models, but we are selective about which one does which job — the lightest capable model handles the routine, high-volume work, and the heavyweight reasoning models are reserved for the small share of tasks that truly need them. The result is that every million tokens of analysis costs about forty grams of CO₂; a full headteacher report, roughly two-and-a-half grams; a Discover answer, under one. And this is our starting point, not our finished one — the figures here include all of our development and testing, and we expect them to fall substantially as the platform matures and we refine it further. As our reach grows, our per-report footprint doesn’t — and that is exactly the point.
Here is the entire footprint of every school Tracker+ has served since September 2025 — all 277.6 million tokens, the whole platform, the whole period, added together. This includes every hour of our own development and testing, so it is if anything an over-count of day-to-day running. Set against everyday reference points, it is genuinely modest.
Measuring honestly is only worth anything if you then act on it. So we fund tree planting through Ecologi — Gold Standard and UK Woodland Carbon Code projects — that over its lifetime will capture more than 600 times the carbon behind our AI. Not offset to the nearest gram, but answered many times over.
Anthropic offers a family of models — from the light, fast and low-energy through to the heavyweight reasoning models. They are not interchangeable in cost or carbon: a larger model can use many times the energy of a smaller one for the same request. So Tracker+ doesn’t send everything to the biggest model available. We match the model to the job, and the numbers show the discipline in action.
Monthly token volume tells the growth story directly — split into input (the prompts, school data and cached context the models read) and output (the analysis they write back). Discover, weekly attendance intelligence and the reports engine have all come online through 2026, and usage has climbed with them. July 2026 set a new record with only half the month elapsed. Crucially, the carbon cost per report has stayed flat throughout — we grow without our footprint growing in step.
There is no official per-token energy figure published by any AI provider, so every number here is a transparent estimate built from peer-reviewed and widely-cited sources. The token counts themselves are exact — taken directly from our usage records for the period. Everything is laid out below so it can be checked or challenged.
Taken from our usage records covering 1 September 2025 to 14 July 2026. We summed every category of input token — whether freshly processed or served from cache — and every output token generated.
We counted all tokens processed, including cached reads, because they still consume real inference compute — this gives the most complete and honest picture rather than a billing-optimised one.
The headline efficiency benchmark is scale-invariant: it divides total carbon by total tokens, so it doesn’t change as volume grows.
Per-report figures apply that blended, real-world rate (caching included — the efficiency the platform genuinely achieves in production) to a representative token footprint for each job type:
Token footprints per job type are representative estimates based on the size of a typical response plus its cached context; the per-token carbon rate is exact from the data. The “mug of tea” reference uses one mug boiled (0.03 kWh × 126 g/kWh ≈ 3.8 g CO₂).
The energy model follows Epoch AI’s analysis of a frontier model, which frames energy as a function of tokens: a baseline query of ~500 output tokens uses ~0.3 Wh, rising to ~2.5 Wh at 10,000 input tokens and ~40 Wh at 100,000 input tokens.
From that we adopt a clearly-stated central rate:
Output is priced higher because each output token requires a full forward pass. These rates sit within the range independently measured on the ML.ENERGY leaderboard (0.15–0.31 J/token for a 32B model), scaled up for frontier-model size and data-centre overhead (H100 GPUs, ~10% utilisation, power-usage-effectiveness included, per Epoch’s assumptions).
Result: 254.4M input × 0.0003 Wh + 23.2M output × 0.0006 Wh = ~90.2 kWh. To be transparent about uncertainty, the plausible range is 45–120 kWh (efficient vs. pessimistic ends of the published figures).
Energy is converted to carbon using the 2025 UK grid average of 126 gCO₂/kWh, reported by Carbon Brief from NESO data (up 2% on 2024’s record-low 124 g).
90.25 kWh × 126 g/kWh = 11,371 g ≈ 11.4 kg CO₂e. Range: 5.7–15.2 kg. Actual figures depend on where and when the inference ran; if Anthropic’s data centres run on cleaner-than-grid power, the true carbon would be lower.
Every comparison is a direct division of the energy or carbon total by a published factor:
Electricity price is the Ofgem July–Sept 2026 cap; car factor is the Energy Saving Trust UK fleet average; home and tree figures use standard UK sustainability-reporting values. All are rounded for readability.
Included: the inference energy of running the models on your tokens, plus data-centre overhead baked into the Epoch figures.
Not included: the one-off energy of training the models (shared across all of Anthropic’s customers, not attributable to Tracker+), the energy of your own servers (Netlify, Supabase), network transfer, or the devices teachers read the reports on. This report scopes deliberately to the AI inference the token data describes.
I wanted to be able to answer a simple question honestly: what does the AI in Tracker+ actually cost the planet? Not a slogan, not a rounded-down guess — the real figure.
So we measured it properly and put the whole thing here, workings and all. If we’re asking schools to trust what we build, the least we can do is show our own numbers first.
Jonathan Parr · Founder
A single snapshot analysis was the right call for the moment, but never the goal. Because we care about getting this right, we’ll keep measuring, keep hunting for efficiencies across every part of Tracker+, and hold ourselves to it year on year. Just as we chart every child’s progress, we’ll chart our own — a public record of an environmental footprint we intend to keep shrinking.