what changed from v1
v1 (feb 2026) leaned on the 2023 luccioni measurements and extrapolation. v2 rebuilds on the 2025–26 disclosures that didn't exist then: google's first per-prompt gemini figures, epoch's and openai's gpt-4o estimates, microsoft's 2026 joule study, and mistral's full lifecycle audit — plus the open-weight frontier (deepseek v4, kimi k3, glm-5.3, mistral large 3) whose architectures are public. the big 2026 story: MoE models kept scaling active params — kimi k3 now fires ~104B/token, as heavy as a closed frontier model, while deepseek's flash tier went the other way to ~16B.
how it works
energy scales with a model's active parameters per token. the coefficient — ~0.007 Wh per 1K output tokens per 1B active params, end-to-end (accelerator + host + idle + cooling, PUE folded in) — is calibrated so a ~100B-active model at a 500-token reply lands on epoch/openai's ~0.3–0.34 Wh gpt-4o figure. CO₂ = energy × grid intensity. water is scaled from google's on-site ~1.1 mL/Wh.
per-model energy & price
| model | architecture | Wh/1K | $/M out |
|---|
| claude haiku | small | 0.12 | 0.50 |
| gemini 3 flash | small | 0.14 | 3.00 |
| claude sonnet | mid | 0.55 | 10 |
| gpt-5 | ~100B active | 0.70 | 10 |
| gemini 3 pro | large | 1.2 | 12 |
| claude opus | frontier | 2.0 | 20 |
| gpt-5 pro | frontier | 2.2 | 120 |
| deepseek v4.1-flash | ~552B / 16B act | 0.11 | 0.60 |
| glm-5.3 | 753B / 40B act | 0.28 | 4.40 |
| mistral large 3 | 675B / 41B act | 0.29 | 1.50 |
| deepseek v4-pro | 1.6T / 49B act | 0.34 | 1.98 |
| kimi k3 | 2.8T / 104B act | 0.73 | 15 |
open-weight MoE models are cheap per token because only a slice of the network fires per token — the energy figure tracks active, not total, params. extended thinking multiplies the token count ~12×, which dominates everything. prices are standard API list, $/1M output tokens, oct 2026 (deepseek shown off-peak, ~2× at peak); note there is no tight link between price and energy — price reflects margin and competition, not watts.
grid intensity · g CO₂e / kWh
clean / market-based ~50 · eu avg ~230 · us avg ~368 · world avg ~480 · coal-heavy ~820.
market-based credits the renewable PPAs hyperscalers buy; location-based counts the actual grid where the data center runs. google's 2025 market-based figure works out to ~125 g/kWh; a coal-heavy region is ~16× that.
no lab publishes audited per-token carbon data, and reported figures span ~2 orders of magnitude depending on accounting boundary (marginal compute vs. whole provisioned fleet). closed-model rows are estimates; open-weight rows use published active-param counts, but specs for 2026 releases (deepseek v4.1, kimi k3, glm-5.3) come from secondary sources and model cards, not papers, so verify before citing. treat every number here as ±2–10×, good for intuition and relative comparison — not a carbon ledger. the honest headline is unchanged from v1: one query is tiny; the aggregate across billions of daily queries is not.