What OpenAI’s Pricing Paradox Means
The pricing paradox of ChatGPT refers to the growing gap between flat subscription fees and the far higher compute costs required to serve heavy users, exposing broken unit economics where more usage can erase margins instead of improving profitability. SemiAnalysis shows how stark that gap is. A USD 200 (approx. RM920) ChatGPT Pro 20x subscription can translate into as much as USD 14,000 (approx. RM64,400) in standard API-equivalent usage if a customer runs the plan to its theoretical limit. That mismatch explains why OpenAI starts losing money on ChatGPT Plus and Pro 5x once utilization passes 11.4%. At the high end, OpenAI’s top tiers move into negative gross margin at only 5.7% utilization, meaning the business model depends on most subscribers using far less than they are allowed on paper.
How Inference Economics Break Subscriptions
The core problem is AI inference pricing: the cost of running large models every time a user sends a prompt. Flat monthly plans were attractive because they made powerful tools feel like consumer software, but they sit on hardware and energy bills that behave more like heavy infrastructure. SemiAnalysis found that token consumption rises sharply with long-horizon coding and agentic tasks, sometimes needing up to 1,000 times more tokens than a simple query. Under current pricing, OpenAI’s margin on ChatGPT Plus and Pro 5x turns negative above 11.4% utilization, while its highest tiers become unprofitable beyond 5.7%. These numbers make clear that premium subscriptions are not priced to sustain frequent, intensive use. Instead, they quietly rely on light or inconsistent users to subsidize power users, a fragile setup as workflows and agents automate more work.
Leaked Financials Show the Scale of the Losses
The strain from these unit economics is visible in OpenAI’s leaked financials. According to financial statements obtained by blogger Ed Zitron and the Financial Times, OpenAI generated USD 13.07 billion (approx. RM60.7 billion) in revenue while total costs reached USD 34 billion (approx. RM157.8 billion), producing an operating loss of USD 20.92 billion (approx. RM97.1 billion). Research and development alone consumed USD 19.18 billion (approx. RM89.0 billion), while sales and marketing climbed to USD 5.73 billion (approx. RM26.6 billion). Yahoo Finance reports that when restructuring and non-cash items are included, OpenAI’s net loss rises to about USD 39 billion (approx. RM181.1 billion). Even though the company reduced its cost-to-revenue ratio from USD 2.37 to USD 1.60 spent per dollar earned, the absolute losses highlight how far current pricing and demand fall short of covering expansion and compute.
Why Premium Tiers Struggle With Agentic Workloads
Premium tiers highlight the subscription profitability gap most clearly. SemiAnalysis estimates that OpenAI begins to lose money on ChatGPT Plus and Pro 5x above 11.4% utilization, a narrow margin that leaves little room for emerging workloads. Agentic systems chain calls, search tools, and code execution, multiplying token use and driving costs much higher than a simple chat interface suggests. SemiAnalysis reports that some agentic tasks can require up to 1,000 times more tokens than a standard prompt. As users push into complex coding, automation, and long-running tasks, a USD 200 (approx. RM920) subscription that can consume USD 14,000 (approx. RM64,400) worth of API-equivalent inference becomes economically impossible at scale. OpenAI’s frontier models remain especially expensive to run, which is why their highest-end capabilities are more likely to be sold via metered APIs than bundled into unlimited-feeling subscriptions.
New Revenue Models: Routing, Partnerships, and Commerce
The widening gap between ChatGPT subscription cost and delivery cost is pushing OpenAI and customers toward new models. Large organizations are already rationing access after seeing bills spike; Microsoft, Meta, and Amazon have reportedly scaled back internal free-for-all deployments, and one company spent USD 500 million (approx. RM2.3 billion) in a single month using Anthropic’s Claude when limits were absent. Many are now routing heavy tasks to cheaper models and reserving frontier systems for high-value queries, a strategy the Wall Street Journal says can cut costs by up to 95%. Others, like Lindy, have moved entirely to alternatives such as DeepSeek V4 to save millions. For OpenAI, this pressure means leaning on usage-based APIs, enterprise partnerships, and commerce-like flows instead of relying on flat consumer subscriptions that cannot support long-horizon, agentic workloads without massive ongoing losses.
