AI token pricing has reached the ROI phase

The AI cost debate is moving from benchmark charts to budget spreadsheets. David Rosenthal’s post argues that major AI platforms grew demand through low subscription prices and generous limits, then began shifting serious users toward token-based billing, tighter rate limits, and more explicit cost recovery. The important question is no longer whether the tools are useful. It is whether their everyday use produces enough measurable return to survive normal enterprise scrutiny.

The post draws on figures and reporting from Ed Zitron and SemiAnalysis, including estimates of how much usage a $200 monthly subscription can consume, OpenAI’s reported 2025 revenue and expense structure, and Microsoft’s move toward usage-sensitive GitHub Copilot billing. Those numbers involve assumptions and should not be treated as audited unit economics for every customer segment. They do, however, explain why the community treated the article as more than a complaint about high prices. The underlying issue is that usage growth eventually has to meet gross margin, procurement policy, and departmental budgets.

The HN thread added useful field context. One commenter described companies that recently moved from broad “use AI” mandates to monitoring, reporting, and escalation around expensive model usage. Another argued that even if models become cheaper quickly, many firms may still cut token budgets if they cannot connect faster code generation or document work to profit. That is the pressure point: productivity anecdotes do not automatically translate into durable willingness to pay.

This matters for product design. Coding agents, long-context assistants, retrieval systems, and document workflows all encourage more inference, not less. If every useful workflow burns more tokens, buyers will push for routing, caching, smaller specialist models, open weights, self-hosting, and stricter access controls. The default model choice becomes a financial architecture decision rather than a purely technical preference.

The article is not proof that AI demand is about to collapse. Hardware, models, and inference stacks can all get cheaper. But it captures a real phase change: the market is beginning to price AI usage as an operating cost. Once that happens, the winning systems are not simply the most capable ones. They are the ones whose capability can be justified repeatedly on an invoice.