Estimated · ML.ENERGY 2025

Claude Water Footprint

Formula v1.3.1 · ML.ENERGY 2025 · ±50%

Drop ~/.claude/water-log.jsonl here

or browse to select the file

Browse file

Zero network requests · No storage · Works fully offline · Close tab to clear data

How we calculate — ML.ENERGY Leaderboard v3.0 · NeurIPS 2025

Formula chain

Water flows through: effective token count → model multiplier → water rate (low / mid / high). All three scenarios are always reported — never just one number.

effective_tokens = (output_tokens × 1.000) // autoregressive decode — formula baseline + (fresh_input_tokens × 0.200) // pricing proxy: $3 / $15 MTok + (cache_creation_tokens × 0.250) // pricing proxy: $3.75 / $15 MTok + (cache_read_tokens × 0.001) // floor value — see note below scaled_tokens = effective_tokens × model_multiplier water_mL = scaled_tokens × rate // three scenarios

Water rates (mL / effective token)

Scenario	Rate	WUE	Basis
Low (optimistic)	0.000036	0.18 L/kWh	AWS modern, cool climate
Mid (central) ✶	0.00036	1.8 L/kWh	ML.ENERGY 2025 US average
High (pessimistic)	0.00079	4.0 L/kWh	Hot climate, older facility

✶ Charts and main values show mid. The low–high band reflects WUE uncertainty across different data centre locations. Anthropic does not disclose which AWS regions handle inference.

Model multipliers (pricing-derived, verified 2026-06-12)

Model	Output price	Multiplier
Haiku 4.5	$5/MTok	0.33×
Sonnet 4.5	$15/MTok	1.00×
Sonnet 4.6 (baseline)	$15/MTok	1.00×
Opus 4.5–4.8	$25/MTok	1.67×
Fable 5 / Mythos 5	$50/MTok	3.33×

Multipliers are derived at calculation time as model output price ÷ baseline ($15). Pricing includes margin, so treat multipliers as ±factor-of-2. Verified against the live Anthropic models documentation on 2026-06-12. Directional accuracy (Haiku < Sonnet < Opus < Fable) holds; absolute magnitudes carry the ±50% uncertainty. Models not in the table fall back to the Sonnet baseline (1.00×).

Empirical anchor (v1.3)

ML.ENERGY Leaderboard v3.0 (Dec 2025): Llama 3.1 70B FP8 on 8×H100 = 0.39 J/output token (GPU, measured via NVML). Apply 1.82× server overhead (GPU ≈ 55% of DGX H100 TDP) → 0.71 J/token at IT load. At US average WUE (1.8 L/kWh): 0.71 ÷ 3,600,000 × 1800 = 0.00036 mL/token (mid rate). Llama 70B is used as the closest publicly-measured proxy for Claude Sonnet — Anthropic does not disclose its hardware or energy use.

Prior anchor (Li et al. 2023, GPT-3 era): 0.014 mL/token — 39× higher. The paper's lead author confirmed (2026-05-31) those estimates should not be applied to modern models. See methodology.md for the full correspondence.

Note on cache_read weight (0.001×)

The pricing proxy gives ~0.02× ($0.30/$15 MTok). At that weight, cache reads dominate long-context Claude Code sessions. The 0.001 floor prevents dominance while acknowledging that some compute is involved. A v2 formula will introduce a cache_size × output_tokens cross-term. The deviation from the pricing proxy is documented in formula.json and absorbed by the ±50% accuracy band.

Verify security

Open DevTools → Network tab → drop your log file → confirm zero requests appear. The page uses no browser storage APIs, no analytics endpoints, and no external resources — view source to verify.