Companies Face Rising AI Costs After Early‑Era Subsidies Fade

AI spending is spiking in 2026 as the “subsidized intelligence” era—where venture capital covered low AI prices—winds down, leaving companies with token bills far higher than expected [1]. Firms such as Uber and Meta report that the surge in usage, especially by AI agents, is outpacing measurable productivity gains, prompting a rethink of how they source and price AI services.

Key takeaways

• Venture‑backed pricing discounts are ending, pushing per‑token rates up to $5‑$30 for new flagship models [1].
• AI agents consume dozens of times more tokens than simple chat interactions, driving higher bills [2].
• Uber exhausted its 2026 AI budget in four months, with little clear link to product improvements [1].
• Meta reversed a policy that rewarded token counts as productivity, warning against usage for its own sake [2].
• Companies are shifting to free, open‑source or domain‑specific models to curb costs, with some small models costing as little as five cents per million tokens [1].

Rising Token Prices and the End of Subsidized Intelligence

The early‑stage AI boom relied on rock‑bottom pricing to lock in users, a strategy described by Kevin Simback of Delphi Labs as “subsidized intelligence” where investors quietly covered the compute bill [1][2]. As OpenAI and Anthropic prepare for public listings, that subsidy is disappearing, and the new rate cards reflect higher per‑token charges—$5 to $30 per million tokens for the latest flagship models versus $2.50 to $15 for the previous generation [1]. Independent analysis shows real‑world costs for the newest models can be 49% to 92% higher depending on prompt length [1].

Agents Amplify Costs and Prompt Corporate Pullback

Unlike chatbots that answer a single query, AI agents perform multi‑step tasks—booking appointments, writing code, moving files—and can spin up dozens of parallel processes, each adding to the token tally [2]. A single agent‑driven task may burn dozens of times more tokens than a standard chat message, and coding workloads are especially token‑intensive [1][2]. This “tokenmaxxing” behavior has led some teams to spend more on tokens than on the employees the tools were meant to augment, a point highlighted by analyst Jack Gold [1][2].

Uber’s chief technology officer disclosed that the ride‑hailing firm blew through its entire AI budget within the first four months of 2026, with roughly 95% of engineers touching AI tools monthly and about 70% of committed code now AI‑generated [1]. Yet Uber’s COO noted the difficulty in tying the rising token spend to concrete product gains [1]. Similarly, Meta’s CTO Andrew Bosworth issued a memo urging staff to stop using AI tools merely to increase token counts, reversing an earlier internal leaderboard that rewarded high usage [2].