Altman publicly admitted OpenAI is losing money on the $200/month ChatGPT Pro plan. Anthropic has been quieter but moved third-party agent tool usage behind a credit meter in May, and Enterprise tier now meters every token at standard API rates on top of seat cost. Axios: "all-you-can-eat" subscriptions may not survive the agent era. The frontier labs are quietly migrating their best customers off flat-rate billing.
PowerPoint is where this lands first. Decks are extraordinarily token-intensive — research, layout reasoning, copy, visuals, template iteration. Writer and Claude lead PowerPoint generation today (Writer especially for enterprise governance and brand-fidelity); GPT-5.5 + Codex is closing the design gap fast and in one reviewer's master-slide test produced a fully themed deck in one shot where Opus 4.7 left the template untouched. ChatGPT Image 2.0 + Codex will likely close the remaining gap within two product cycles.
For a CAIO: token economics now belongs alongside vendor risk in every AI deployment plan. Three questions per tool launch: which model handles which task (Haiku vs. Opus is a 5–25× cost swing), what gets cached vs. regenerated, and how usage gets attributed. A single autonomous agent running all day can cost $1,000–$5,000 under metered billing — a different conversation than a $20 seat.