Palantir CEO Alex Karp on Wednesday called the per-token pricing model used by OpenAI and Anthropic fundamentally broken. “Something has gone completely wrong,” Karp said, according to CNBC.
His core argument: escalating token costs are forcing enterprise customers to choose open-weight models over proprietary closed ones, and to prioritize inference efficiency over aggressive token consumption.
The Enterprise Math Problem
Per-token pricing works when usage is predictable. A chatbot answering customer questions generates a roughly stable number of tokens per session. The economics change when autonomous agents enter the picture.
Agents consume tokens non-linearly. A single task execution can involve dozens of API calls: reasoning loops, tool-use chains, retry logic, context window management, and multi-step planning. Each step burns tokens. At scale, with hundreds or thousands of agents running concurrently, the bill compounds in ways that per-token pricing was never designed to handle.
Karp’s criticism, reported by CNBC, signals that enterprise customers are not just complaining about costs in private. A CEO running a $80+ billion market cap company is saying it publicly.
The Open-Weight Migration
The logical response to unsustainable token costs is to move workloads to models where inference is a fixed infrastructure cost rather than a metered API charge. That means Llama, Qwen, Mistral, and the growing ecosystem of open-weight models that companies can deploy on their own hardware or through compute providers.
Coinbase demonstrated this path last week, cutting AI infrastructure costs by 50% after routing workloads to cheaper models (including Chinese open-weight alternatives like GLM 5.2 and Kimi 2.7) and increasing caching hit rates from 5% to 60%. The Coinbase case validates Karp’s thesis with concrete numbers.
Pricing Pressure Ahead
If large enterprises follow the pattern Karp describes, OpenAI and Anthropic face a pricing dilemma. Lower per-token prices to retain customers, and margins compress on a business model that already burns cash. Maintain current pricing, and customers migrate workloads to open alternatives where they control the cost structure.
The timing of Karp’s public critique matters. It arrives the same day Meta announced a cloud compute business to sell excess AI capacity, which would add another source of inference supply to the market. More supply, combined with enterprise pushback on pricing, creates downward pressure on what model providers can charge.
For teams deploying agents in production, the signal is clear: the economics of which model runs your agents are becoming as important as the model’s capability. Token costs at scale are no longer a rounding error.