Backside line: The maths behind AI subscriptions is beginning to look uncomfortable. Flat month-to-month pricing helped gasoline the fast adoption of instruments like ChatGPT and Claude, however new evaluation suggests these charges might not come near overlaying the precise price of heavy use. As customers push these programs tougher and extra demanding AI workflows take maintain, the hole between income and compute prices is changing into troublesome to disregard.
SemiAnalysis has calculated how massive that hole actually is. After testing subscription tiers from each OpenAI and Anthropic – operating long-horizon coding and agentic duties till weekly limits have been exhausted – the agency discovered that the price of theoretical most utilization of those plans if priced at normal API charges far exceeds what customers really pay.
A $200 ChatGPT Professional 20x subscription might price as a lot as $14,000 in API pricing if totally utilized. Anthropic’s Claude Max 20x plan, additionally priced at $200 per thirty days, has a comparable ceiling, with potential utilization totaling roughly $8,000 in token prices.
These figures assist clarify why utilization charges matter a lot to the AI firms providing them. In response to SemiAnalysis, Anthropic breaks even on Claude Professional and Claude Max 5x at round 20% utilization. OpenAI’s margin is thinner. It begins shedding cash on ChatGPT Plus and ChatGPT Professional 5x as soon as utilization climbs above 11.4%.
– SemiAnalysis (@SemiAnalysis_) June 10, 2026
The economics get tighter on the excessive finish. Anthropic reaches zero gross margin at roughly 10% utilization on its top-tier plans, whereas OpenAI crosses into adverse territory at simply 5.7%. It does not take excessive use for these subscriptions to show unprofitable.
Adjusting pricing or proscribing entry will not be an easy repair. Subscription fashions have been central to consumer development, and pulling again dangers slowing momentum in a market the place capabilities stay a key aggressive differentiator.
A part of the stress comes from how AI is definitely getting used. Token consumption is rising rapidly, particularly with agentic programs that may require as much as 1,000 occasions extra tokens than a regular immediate. That form of demand is already forcing massive organizations to rethink how freely these instruments needs to be deployed.
Microsoft, Meta, and Amazon have reportedly pulled again from inside efforts that inspired heavy utilization after prices escalated. In a single broadly cited instance, an organization burned by way of $500 million in a single month utilizing Anthropic’s Claude, largely as a result of it did not put limits on worker entry.
That form of overspending is pushing firms towards extra managed approaches. One technique gaining traction is to shift workloads between fashions relying on the duty. Extra advanced queries go to costly frontier fashions, whereas routine work is dealt with by cheaper alternate options.
– Dong Ming (@dming) June 14, 2026
The financial savings might be substantial. A Wall Avenue Journal report discovered that routing duties this fashion can minimize prices by as much as 95%. “You do not want a mannequin that is aware of quantum gravity,” Columbia College vice dean Vishal Misra advised the publication. “These open-source fashions are very succesful, and the flexibility to cost a giant premium for AI goes to decrease.”
Some firms have already made the shift. Flo Crivello, founder and CEO of AI assistant startup Lindy, introduced that the corporate moved 100% of its visitors to DeepSeek V4, switching totally away from Anthropic’s fashions. DeepSeek V4 proved similar to Claude Sonnet at a fraction of the associated fee, and the transfer has “saved the corporate thousands and thousands of {dollars},” Crivello mentioned.
Others are going additional by constructing their very own AI programs on prime of open-source fashions educated on inside information. Whereas that requires extra upfront funding, it provides tighter price management and reduces dependence on third-party suppliers. In some instances, these tailor-made programs might even outperform general-purpose frontier fashions for particular use instances.
There’s some expectation that prices will ease over time. As infrastructure expands and newer fashions exchange older ones, the price of operating mid-tier programs ought to decline. SemiAnalysis means that fashions on the Opus 4.8 stage might finally be delivered profitably for round $20 per thirty days.
That doesn’t apply to probably the most superior programs, although. Frontier fashions, together with these nonetheless in growth, stay costly to run. Their highest-end capabilities might more and more be priced through APIs reasonably than bundled into shopper subscriptions.
For now, AI suppliers are juggling two forces: customers need highly effective instruments at low, predictable month-to-month costs, however the infrastructure to run them stays pricey and extremely delicate to utilization. OpenAI CEO Sam Altman has acknowledged the strain, noting that rising token prices have gotten a critical difficulty and that the corporate is working to assist customers “get extra worth for much less spend” when utilizing ChatGPT.
