Anthropic analyzed 207,049 privacy-preserving Claude and Claude Code research sessions (August–September 2026) classified via the 7,262-task SciNet taxonomy to quantify AI adoption across scientific fields and tasks. Formal sciences—computer science, mathematics, statistics—show disproportionately high usage relative to researcher population shares; computer science alone accounts for 32.7% of Claude Code sessions despite representing 4.3% of global researchers, while computational subfields (computational biology, linguistics, neuroscience) and robotics are strong adopters. Conversely, lab-intensive physical sciences (chemistry) and patient-facing medicine show low relative usage. By task, data analysis dominates (33.7% of Claude, 48.1% of Claude Code sessions vs. 27.6% of researcher time), followed by writing/communication (32.0% on Claude vs. 16.0% time share). Data gathering (28.6% time, 11.5–11.9% sessions) and ideation/hypothesis generation (7.1% time, 2.1–4.2% sessions) are underrepresented, though a survey of 528 AI-active researchers ranks these as the two most important tasks for discovery. Theoretical AI capability scores, derived via Bradley–Terry modeling over task steps, strongly predict usage: top-20% capability tasks receive 2× (Claude) to 4× (Claude Code) expected use; bottom-20% tasks (e.g., purifying reaction mixtures) see minimal use. Capability correlates with reinforcement-learning suitability (r=0.78), cognitive work (r=0.69), and negatively with physical (−0.57), interpersonal (−0.31), and regulatory (−0.38) components. Specialized Claude Science sessions in biology/biomedicine show longer, more autonomous (70%+ high autonomy vs. 23.7% Claude, 64.1% Code), doctoral-level (80%+ vs. 50–70%) work, with ideation reaching 8% of sessions. Modeling scenarios suggest current AI could double output in some fields; full non-physical automation could yield 15× gains in computer science/statistics versus 3–4× in bench sciences constrained by physical bottlenecks. About 40% of surveyed researchers expect AI to execute full projects autonomously by 2030.