Anthropic Claude Science Ships Amid Flawed Carbon Accounting in Offsets

Anthropic announced Claude Science at a pharmaceutical executive event, extending the Claude Code architecture to drug development workflows and internal rare disease research. The model integrates specialized biology tool use and high-level task autonomy. US export controls on Mythos and Fable were simultaneously lifted after extended negotiations.

California's Low Carbon Fuel Standard manure-to-RNG credits pay farmers $100-200 per tCO2e avoided, yet 2023-2025 field data from 120 dairies show baseline methane emissions 35% lower than modeled, inflating credits. Parallel AI training runs for models like Claude 3.5 consume 200-400 MWh per 10^25 FLOP, producing 80-160 tCO2e that current voluntary offsets frequently source from the same inflated agricultural pools.

The pattern reveals systemic mismatch: AI deployment metrics track parameter counts and benchmark scores while carbon accounting relies on unverified counterfactuals identical to those rejected in CARB's 2025 program review. Operational cost of inference therefore remains externalized rather than internalized as direct emissions liability.

Next reporting cycle will test whether Anthropic publishes per-training-run Scope 3 figures against verified additionality standards or continues reliance on unadjusted offset vintages through 2027.