AI usage report · July 2025 to June 2026
Sample. The data is fictional; the format is what the delivered report looks like.
Claude usage footprint
Methodological range: 1.14 t to 2.84 t of CO₂e.
of CO₂e over the period, for 12,573 Mtok of processed tokens.
That is roughly 15,199 km by car and 492 showers.
Volumes measured server-side by Anthropic.
Spend
At Anthropic's public prices, cache included. Spend doubles as the internal reference: the same token volumes drive both cost and footprint. Your actual invoice may differ if you have negotiated pricing; the footprint only depends on volumes.
89% of the period's tokens are cache re-reads, counted at the reduced factor published in the methodology.
Model mix
CO₂e per month and per model family. The mix is your main lever: for the same usage, model choice moves both cost and footprint.
| Model | Input | Output | Cost | CO₂e | Confidence |
|---|---|---|---|---|---|
| Claude Sonnet | 8,670 Mtok | 694 Mtok | 59,533 € | 1.41 t | direct measurement or calibrated decomposition |
| Claude Opus | 633 Mtok | 51 Mtok | 7,530 € | 210.4 kg | direct measurement or calibrated decomposition |
| Claude Haiku | 2,278 Mtok | 247 Mtok | 4,920 € | 205.7 kg | direct measurement or calibrated decomposition |
57% of these tokens go through Claude Code, the rest through the direct API.
Intensity
Footprint relative to output rather than raw volume: the reading that lets you compare two periods or two teams without penalising the one that ships more.
Scope: these counters come from Claude Code analytics (no one is named, only totals are kept). Direct API usage is included in the footprint but not in the output counters.
Routing decisions worth studying
Two volume shifts to the model family one step down, priced at constant volume over the period. Orders of magnitude to support a decision, not commitments.
Route half of the Opus volume to Sonnet
Long orchestration tasks stay on Opus. The rest moves to Sonnet, one capability step down. Computed at constant volume over the period.
Route a fifth of the Sonnet volume to Haiku
Short reviews, rewrites, extraction: Haiku's capability is enough there. Computed at constant volume over the period.
Month by month
| Month | Tokens | Cost | Energy | CO₂e | Water |
|---|---|---|---|---|---|
| July 2025 | 547 Mtok | 3,309 € | 222.8 kWh | 82.7 kg | 1.3 m³ |
| August 2025 | 602 Mtok | 3,637 € | 244.4 kWh | 90.7 kg | 1.5 m³ |
| September 2025 | 800 Mtok | 4,824 € | 324.2 kWh | 120.3 kg | 1.9 m³ |
| October 2025 | 945 Mtok | 5,725 € | 384.5 kWh | 142.7 kg | 2.3 m³ |
| November 2025 | 1,039 Mtok | 6,289 € | 421.3 kWh | 156.4 kg | 2.5 m³ |
| December 2025 | 892 Mtok | 5,378 € | 360.1 kWh | 133.7 kg | 2.2 m³ |
| January 2026 | 1,133 Mtok | 6,869 € | 460.4 kWh | 170.9 kg | 2.8 m³ |
| February 2026 | 1,171 Mtok | 6,576 € | 451.2 kWh | 167.5 kg | 2.7 m³ |
| March 2026 | 1,262 Mtok | 6,905 € | 478.9 kWh | 177.8 kg | 2.9 m³ |
| April 2026 | 1,329 Mtok | 7,172 € | 500.0 kWh | 185.6 kg | 3.0 m³ |
| May 2026 | 1,396 Mtok | 7,495 € | 522.4 kWh | 193.9 kg | 3.1 m³ |
| June 2026 | 1,458 Mtok | 7,805 € | 543.6 kWh | 201.8 kg | 3.3 m³ |
| Total | 12,573 Mtok | 71,982 € | 4.9 MWh | 1.82 t | 29.5 m³ |
Scope and exclusions
- Measurement source
- Token volumes come from the Anthropic Admin API, which measures them server-side, model by model. Nothing is estimated client-side, nothing is sampled.
- Claude Code seats
- Claude Code seat usage (subscriptions) was queried and is included in this measurement, alongside API usage.
- Cache treatment
- Cache writes count as full input, for energy and for price. Cache reads count at 8% of the energy of an input token (the memory re-read residual documented in the methodology) and at the public price of 0.1x input. Tokens read from cache are shown separately, outside the processed-tokens counter.
- Period boundaries
- The period covers July 2025 to June 2026, in calendar months: from the first day of the first month at 00:00 UTC to the first day of the month after the end, exclusive. Every boundary is set in UTC, never in local time.
- Factor coverage
- 100% of the period's tokens are covered by published emission factors: every gram shown traces back to a documented model.
- What this document is
- This report measures the footprint of a usage: Claude model inference, datacenter share. It is not an organisational GHG assessment: end-user devices, networks and model training remain out of scope, as documented in the methodology.
Methodology annex
Token volumes are measured server-side by Anthropic, not estimated. Each model is converted to energy, CO₂e and water using published, versioned factors, then the infrastructure parameters below apply. Full equations live at tokenclimate.com/methodology.
| PUE | 1.14 |
| Grid carbon intensity (CIF) | 0.287 kgCO₂e/kWh |
| On-site water (WUE) | 0.18 L/kWh |
| Off-site water (EWIF) | 5.11 L/kWh |
| Amortised manufacturing | 44 gCO₂e/kWh |
Uncertainty range
The total lies between 1.14 t and 2.84 t of CO₂e according to the published ranges of the methodology (factors tokenclimate-v3-2026-07): cache read factor 0.05 to 0.15, grid carbon intensity ±30% depending on location, amortised manufacturing 22 to 66 gCO₂e/kWh. Parameters without a published range stay at their central value.
Conversions and equivalences
Costs are converted from Anthropic's public USD prices into euros at the ECB reference rate of 22 June 2026: 1 USD = €0.8729. Equivalences: 120 gCO₂e per kilometre for an average European passenger car (ADEME factor, well-to-wheel) and a 60-litre shower (ADEME).
Wording ready to paste into your responses and reports
“Over the period July 2025 to June 2026, Claude (Anthropic) usage amounts to 1.82 t of CO₂e for 12,573 Mtok of processed tokens (cache excluded). Volumes are measured server-side by the provider. Conversion follows the public TokenClimate methodology (factors tokenclimate-v3-2026-07, tokenclimate.com/methodology), versioned and sourced.”
Glossary
- Mtok
- One million tokens, the volume unit used throughout this document.
- Input / output tokens
- Input is the text sent to the model (prompt, context), output the text it generates. Output costs far more energy per token than input.
- Cache
- Reuse of context that was already processed. A write stores the context, a read reuses it at reduced cost and energy.
- CO₂e
- Carbon dioxide equivalent: the unit that aggregates greenhouse gases by their warming potential.
- Wh / kWh / MWh
- Watt-hour and its multiples. Electrical energy consumed by the inference servers, before the PUE is applied.
- T1 / T2 / T3
- Source tiers of the parameters: T1 measurement or published primary report, T2 documented derivation from T1 sources, T3 engineering estimate.
References
- Jegham N., Abdelatti M., Hendawi A., "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference", arXiv:2505.09598, v6, 2025.
- Mistral AI, "Our contribution to a global environmental standard for AI", 2025.
- Google, "Measuring the Environmental Impact of Delivering AI at Google Scale", arXiv:2508.15734, 2025.
- Luccioni A. S., Jernite Y., Strubell E., "Power Hungry Processing: Watts Driving the Cost of AI Deployment?", ACM FAccT 2024.
- "EcoLogits: Evaluating the Environmental Impacts of Generative AI", Journal of Open Source Software, 2025, doi:10.21105/joss.07471.
- NVIDIA, "Product Carbon Footprint Summary: HGX H100", 2024; Boavizta, BoaviztAPI.
- AFNOR, SPEC 2314, "Référentiel général pour l'IA frugale", 2024.
- Llopis P., arXiv:2606.10660, 2026.