Kai ChenAnthropic has committed $150 million to fellowships while OpenAI is reportedly weighing token-price cuts that neither lab has posted. One is a concrete outlay; the other is still a market signal. They belong to the same bet.
Cheaper output rewards systems that consume more of it
OpenAI is reportedly weighing drastic token-price cuts in anticipation of cuts from Anthropic. Neither company has announced a change, but the planning assumption matters: OpenAI is preparing for raw model access to get cheaper.
No one needs to coordinate this. Competition does the coordinating. Once one supplier can lower the price of an input, every rival has to consider the same move or explain why its output deserves a premium.
That changes inference economics beyond the API invoice. When model calls are expensive, applications minimize them and labs can capture value at the point of access. As calls get cheaper, it becomes rational to build systems that use more of them: agents that plan, retry, invoke tools, transact, and remain inside a workflow until the task is complete.
Margin compression does not imply usage compression. The unit can get cheaper while the system consumes far more units. The valuable layer becomes the one that turns abundant inference into completed work.
OpenAI is buying the path from prompt to payment
OpenAI’s acquisition of agent-cloud provider Ona for Codex and its partnership with Visa on agent payments are two versions of the same move. One adds an execution environment around a coding agent. The other adds a transaction rail for agents that need to pay.
Seen separately, these look like product expansion and business development. Seen together, they describe a new unit of competition. The model call begins the process; the agent environment carries it forward; the payment connection lets it cross from recommendation into action.
A token does not own the checkout.
As inference becomes a cheaper component, value shifts toward orchestration, distribution, and task completion. Acquiring Ona gives OpenAI more control over where Codex agents operate. Working with Visa extends the agent strategy toward transactions. Both moves make more sense if the lab expects the intelligence underneath them to face pricing pressure.
The acquisition and partnership do not prove that a price cut is settled. They show that OpenAI is building a business architecture that does not require expensive tokens to remain the main source of differentiation.
Anthropic is securing institutions and capacity
Anthropic is making the same structural adjustment through different complements. Its fellowship program expands Claude’s institutional footprint, while new direct data-center leases place the company closer to the capacity supporting its models.
The fellowship program works on adoption and affiliation. The leases work on supply. Together, they move Anthropic beyond selling access to Claude and toward controlling more of the conditions under which Claude is developed, supported, and embedded.
OpenAI is assembling execution and payment channels around Codex. Anthropic is deepening institutional reach and its direct relationship with infrastructure. The assets differ because the companies occupy different positions, but the incentive is shared: if token margins narrow, durable value has to sit elsewhere in the stack.
That does not make the underlying models unimportant. It makes them foundational in the literal sense. Foundations are essential, expensive, and rarely where the occupants spend most of their time.
The price signal is real before the price cut is
The reported OpenAI cuts and anticipated Anthropic cuts are not posted pricing changes. Competitive preparation is evidence of an expectation, not a receipt.
Price is also not the only gate. Access policy can constrain model usage even when the nominal cost falls. Cheaper tokens only commoditize the uses that suppliers permit and customers can actually move between. The margin shift therefore depends on competition, accessible models, and enough customer mobility to make price matter.
The absence of a posted cut sharpens the pattern rather than dissolving it. Over the same arc in which token prices became a competitive variable, both labs invested in complements that retain value under cheaper inference: agent clouds, payment channels, institutional programs, and direct capacity relationships.
The $150 million commitment and the unposted price signal are two sides of the same wager. One puts capital around the model; the other marks pressure on the model’s unit price. The token remains the foundation. The margin moves upstairs.