The Era of AI Yield (Capgemini, October 2026): seven engineering levers cut token spend 40% to 60%

Capgemini Invent's first AI Compass issue is a 60-page guide to token economics. Its starting point is that free and below-cost model access is ending: providers are moving to metered billing, usage caps and peak-hour limits, so inference becomes an industrial input cost. Citing the Atlanta Fed, it notes US firms spent $1,358 per employee on AI in 2025 and expect $2,068 in 2026. The report lays out eight convictions and seven engineering levers, including caching that can cut API calls by up to 68.8%, routing routine work to open-weight models in the 25 to 35 billion parameter range that cost one to two orders of magnitude less per token, selective reasoning that cut tokens 48.5% while improving accuracy, and tool-call quotas for agents. Applied together, Capgemini documents 40% to 60% reductions in token spend at similar output quality. For founders, the warning is about pricing: a product sold per seat with a cost base that scales per token carries unhedged exposure to usage growth. Tokens per outcome, not tokens consumed, is the metric to manage.