Empirical (Tier 1) = API benchmark data from Jegham et al. 2025 (±50%). Estimated (Tier 3) = order-of-magnitude (±1000%).
Energy is facility-level (IT equipment × datacenter PUE).
Carbon uses the selected grid region (default: global average 436 gCO₂e/kWh · IEA 2023).
Regional values are static annual averages last updated January 2024 — not live.
Grid intensity changes as renewables penetration shifts; treat these as indicative.
Water uses a global default of 1.7 L/kWh regardless of region (regional WUE supported in the SDK only).
Why might these numbers differ from EcoLogits or other tools?
Prompt-length sensitivity. Jegham et al. measured energy separately for short (<1k tokens), medium (1k–5k), and long (>5k) prompts. Short prompts cost more per token because fixed per-request overhead is spread across fewer tokens. Some tools use a single average value; this calculator uses the matching measured bucket.
Provider-specific PUE. Datacenter overhead is applied using official sustainability reports: Azure 1.12, AWS 1.14, Google 1.10. Tools that use a global average (typically 1.2–1.3) will diverge, particularly for Google and Anthropic models.
Grid carbon intensity. Regional values are static annual averages sourced from IEA, Ember, and regional grid operators, last updated January 2024. Three important caveats: (1) grid intensity changes year-on-year as renewables penetration grows — figures may be materially different from current actuals; (2) these are grid averages, not provider-adjusted — Google, Microsoft, and AWS purchase renewable energy certificates (RECs) and power purchase agreements (PPAs) that may reduce their effective intensity below the regional average; (3) providers route requests dynamically and do not expose per-request datacenter location, so the selected region is your assumption, not a guarantee. For real-time grid intensity, use the vetch SDK: set ELECTRICITY_MAPS_API_KEY=your_key and VETCH_REGION=us-east-1. A commercial Electricity Maps plan is required for production use.
Water scope. Water figures here reflect operational datacenter cooling only. Tools that include chip manufacturing and lifecycle water will show higher values.
These numbers represent our best current understanding — they are not guaranteed to be correct.
Full methodology: PROVENANCE.md.
Something look off? Email marco@prismaticlabs.ai or open an issue.