how it works
carbon = energy per token × grid intensity × data center PUE (~1.2). token-level energy from luccioni et al. (2023), who instrumented bloom-176b and opt-175b directly. frontier estimates (opus, gpt-4) extrapolated from model-size scaling.
per-model baselines · location-based · us grid
haiku / small (~7–20B): 0.04 g CO₂e / 1K tok
sonnet / mid (~30–70B): 0.2 g CO₂e / 1K tok
opus / gpt-4 (frontier ~500B+): 2.0 g CO₂e / 1K tok
reasoning (extended thinking): 6.0 g CO₂e / 1K tok
market-based divides by ~6, reflecting renewable energy purchasing by anthropic, google, openai.
no ai company publishes per-token carbon data. these estimates carry ±50–200% uncertainty. hardware generation (a100 vs h100), batch size, quantization, and location all materially affect real emissions.