Palantir CEO Alex Karp attacks token-based AI pricing used by OpenAI and Anthropic, says “something has gone completely wrong”

Palantir CEO Alex Karp used a recent appearance to criticize the token-based pricing model used by AI providers such as OpenAI and Anthropic, arguing that the approach has created a dysfunctional outcome for enterprise customers. In comments reported by Yahoo Finance, Karp said, “Something has gone completely wrong,” and linked the problem to rising costs that, in his view, are changing how businesses evaluate AI spending.

Token-based pricing is a billing method where customers pay based on the amount of data processed by a model, commonly measured in “tokens,” which roughly correspond to chunks of text. Karp’s argument, as summarized in the market coverage, is that this structure can make costs hard to predict and can erode return-on-investment for companies trying to deploy AI at scale.

Multiple outlets reported that Karp characterized enterprises as increasingly frustrated or “empty-handed” under token-led economics, especially as usage grows and bills rise. Quartz described the criticism in terms of customer frustration around the token approach, while Benzinga also framed Karp’s remarks around a shift toward models aimed at lower cost and more direct alignment with enterprise ROI goals.

Karp did not limit his remarks to critique. CNBC reported that he argued Palantir and Nvidia’s open model strategy offers a more workable path for businesses looking for cost-effective deployment. In that framing, “open-weight” models, where the model weights are available rather than fully closed behind a proprietary API, are positioned as a way to control costs and improve deployment flexibility compared with traditional token-metered access.

The comments land as AI competition increasingly extends beyond model quality into pricing, deployability, and operational economics. For enterprise buyers, the core question is less whether an AI system can produce answers and more whether it can be deployed reliably, integrated with business workflows, and run within budgets as usage expands.

Palantir’s own business is tied to helping organizations deploy AI and data analytics in operational settings, and Karp’s public critique of token pricing reads as consistent with a broader message: enterprise adoption depends on total cost and predictable economics, not only model performance. By pushing open-weight alternatives, he is effectively drawing a contrast between API-driven consumption and deployment approaches that can be optimized for specific workloads.

Still, the criticism leaves several details unclear. The market reporting summarized Karp’s thesis and included his headline quote, but it did not provide specific pricing figures, customer contract terms, or quantified performance-and-cost comparisons between Palantir’s approach and the token models he criticized. Without those numbers in the public account of the remarks, it is difficult to assess how widely his claims apply across industries and use cases.