Palantir Technologies CEO Alex Karp leveled sharp criticism at OpenAI and Anthropic, arguing that the token pricing models of major AI companies have broken down and pushed enterprises toward alternative solutions.
Karp contends that escalating token costs have created a fundamental market dysfunction. Rather than paying premium prices for proprietary models, companies now face economic pressure to adopt open-weight alternatives where they can control computational efficiency and avoid what he calls "tokenmaxxing," the practice of inflating token counts to increase API revenue.
The critique strikes at a core revenue driver for OpenAI and Anthropic. Both firms price access to their large language models on a per-token basis. GPT-4o charges $15 per million input tokens and $60 per million output tokens. Claude 3.5 Sonnet costs $3 per million input tokens and $15 per million output tokens. As enterprises deploy AI at scale, these costs compound rapidly. A single complex query can consume thousands of tokens, making API expenses unpredictable and substantial.
Karp's argument aligns with observable market trends. Open-weight models from Meta (Llama), Mistral, and others have gained traction among cost-conscious enterprises. These models run on in-house infrastructure, eliminating per-token fees and providing transparency into computational expenses. Companies can also fine-tune open models for specific tasks, reducing token consumption through optimization.
Palantir itself has positioned its data platform as infrastructure for handling AI workloads efficiently. The company offers tools for enterprises to manage large language models internally rather than relying exclusively on third-party APIs. This positions Palantir to benefit from corporate migration away from token-based pricing.
The broader implication concerns whether OpenAI and Anthropic can sustain current pricing power as competition intensifies. If enterprises systematically shift toward open-weight alternatives, revenue growth at API-dependent AI firms could decelerate. OpenAI has already acknowledged competition from open models in internal documents. Anthropic faces similar pressures as its Claude models gain adoption but face price resistance from budget-constrained buyers.
Karp's comments reflect real economic constraints facing enterprise AI spending. Companies seeking to deploy AI responsibly need predictable costs and control over efficiency metrics. Token-based pricing obscures both.
Investors tracking AI infrastructure and API economics should monitor GPT-4 adoption rates and benchmark them against Llama and open-source alternatives in enterprise deployments. Watch OpenAI and Anthropic's pricing announcements, token efficiency improvements, and revenue guidance for signs of margin compression.