Anthropic cuts off volume discounts the moment an enterprise burns through its purchased tokens and makes customers renegotiate from scratch, The Information reports, while OpenAI is offering more flexible terms alongside price cuts of 50% to 80% on its cheaper models. Read the gap correctly: one vendor is rationing scarce compute, the other is discounting to fill it.
Anthropic's enterprise contracts have a cliff edge. Buy a block of tokens at a discounted rate, use them up, and the discount disappears with them, forcing a fresh negotiation rather than a rollover, The Information reports. OpenAI is competing for the same procurement budgets with flexible terms and price cuts, including an 80% cut to its Luna model in July and new GPT-6 Sol and Luna models priced 50% below their predecessors. The difference is explained by capacity. Anthropic's revenue run-rate went from roughly $9 billion at the end of 2025 to more than $65 billion by the end of July and is now pacing above $100 billion annualized. A company selling that much faster than it can add compute has little reason to discount. It has to ration.
The contract's logic is simple. A discount moves inventory when supply exceeds demand. When demand is greater than supply, the seller allocates scarce capacity to customers willing to pay full price and lets others wait or leave. Anthropic's token cliff does that. A customer that uses its contracted volume has shown that Anthropic's compute is worth more to it than the locked-in price, giving Anthropic an opportunity to reprice at renegotiation. Capacity-constrained sellers use the same logic when the queue of full-price buyers is long.
The same industry offers a recent precedent. When Nvidia's H100 GPUs became the bottleneck for the AI buildout in 2023, Nvidia kept prices firm, allocated chips to customers it judged most strategic and left others with lead times of most of a year. AMD pitched its MI300 line as the available alternative. Nvidia's terms loosened only after its own supply and its competitors' credibility caught up, which took the better part of two years. Anthropic is applying that scarcity model to tokens instead of silicon.
OpenAI's flexibility looks like weakness. It has paused training twice in three months after a string of rogue-agent incidents, and Micah Carroll, its RSI preparedness lead, said after Friday's incident report that "all inference for our most capable models remains stopped until we have hardened our systems further." A vendor that cannot reliably run its top-tier model has a weaker case for premium pricing. OpenAI is stacking 50% to 80% price cuts with generous renegotiation terms because it needs enterprises to keep signing while Anthropic is constrained and distracted by its own growth.
Capacity explains the gap, and it is closing on Anthropic's own timeline. The company signed an $11.6 billion cloud deal with Akamai last week, the latest step in buying compute its revenue has outrun. Anthropic is buying its way out of a shortage. Once that capacity lands, the incentive to ration by price weakens because rationing works only while the queue of full-price buyers is longer than the available compute. The desk expects Anthropic's discount cliff to soften as capacity comes online, driven by supply rather than enterprise pressure or bad press.
Enterprises that feel punished by a vendor can leave it, and switching between large language model APIs is supposed to be close to frictionless: the request-response pattern is similar, prompt formats overlap, and a procurement team can point a workload at a new endpoint in an afternoon. That could send customers to OpenAI faster than the capacity argument suggests. Production switching costs are higher than the API call, however. They include the eval suite, the security and compliance review, fine-tuned prompts and guardrails built around one model's quirks and, for regulated buyers, months of vendor-risk review. Those pieces do not transfer cleanly or cheaply. Anthropic is also selecting for customers least likely to churn over price: the customers burning through contracted tokens fastest are the ones most dependent on Claude, which is the group a capacity-constrained seller wants to keep.
A political risk sits alongside the pricing story. A memo attacking Dario Amodei's record and Democratic ties circulated inside the White House ahead of his Sunday dinner with Trump, the second such file aimed at an AI executive in recent months. Axios reported the administration is trying to keep everyone, including Amodei, "seated at the table," and Trump has since confirmed the meeting while repeating his dismissal of AI safety fears. That adds a modest risk premium to the current front-runner and reminds investors that Washington's reach can outlast commercial momentum.
Buyers negotiating with Anthropic should treat the token cliff as a known contract term. They can push for rollover of unused tokens and a true-up clause that locks in the discounted rate for a defined renegotiation window before signing. That leverage exists while Anthropic's compute remains constrained and may disappear once supply catches up with demand. The desk does not expect Anthropic to loosen these terms before its new capacity comes online, which on current announcements means 2027. OpenAI's generosity lasts until its most capable models are running again.