Palantir CEO Alex Karp has publicly criticized the token-based business model used by OpenAI and Anthropic, arguing that the way major model providers charge for AI usage reflects a deeper problem in the market. The remarks, reported by CNBC, are notable because they come from the head of a company that has been pushing a different pitch to enterprise and government customers: AI as an operational system tied to outcomes, not just model access metered by tokens.

The immediate news is less about a product launch than a widening strategic divide in AI. On one side are model companies such as OpenAI and Anthropic, whose APIs and chat products are commonly priced by usage. On the other are platform vendors like Palantir that are trying to position themselves as the layer that connects models to sensitive data, business workflows, and deployment controls. Karp’s criticism matters because it captures a tension many buyers already face: whether enterprise AI spending should center on raw model consumption or on software systems that manage reliability, cost, security, and workflow integration.

CNBC’s available excerpt does not include Karp’s full comments, the venue where he made them, or detailed examples supporting his critique. That limits how far the reporting can go on his specific argument. But even from the headline alone, the thrust is clear: Palantir is challenging not just competitors, but the economics that have become standard across much of generative AI.

Why Karp’s criticism matters now

Karp’s attack lands at a time when enterprise buyers are moving past early experimentation with large language models and asking harder questions about cost predictability. Token pricing has helped companies like OpenAI and Anthropic scale access to frontier models because it maps usage to compute demand. For developers, it is a straightforward way to buy model capacity through an API. For finance teams and CIOs, though, token metering can be difficult to forecast once AI features spread across multiple products, teams, and agents.

That is where Palantir has tried to differentiate. Rather than competing head-on as a foundation model company, Palantir has framed its value around orchestration, governance, and deployment inside complex institutions. Its broader enterprise AI narrative has been tied to platforms such as AIP, short for Artificial Intelligence Platform, which Palantir has marketed as a way to operationalize models in real-world settings with controls around data access and decision workflows.

Karp’s comments therefore read as more than a rhetorical swipe. They align with Palantir’s long-running argument that enterprise AI cannot be reduced to access to a powerful model. In sectors with large datasets, regulatory constraints, or high-risk decision chains, the pricing model is inseparable from system design. If the base economics reward constant token consumption, buyers may worry that software architecture is being shaped around usage expansion rather than efficiency.

The business model fight behind the model race

The pricing debate is really a fight over where value accrues in the AI stack. OpenAI and Anthropic sit close to the model layer, where performance improvements and new capabilities can justify premium pricing. Token-based charging is a natural fit for that business because it turns model calls into recurring revenue. It also lets vendors segment offerings by model quality, context length, and throughput.

Palantir comes from a different software tradition. Its enterprise and government contracts have historically emphasized deployment, integration, and mission-specific software. From that perspective, usage-based model fees can look like an unstable foundation for long-term enterprise AI budgets, especially when workloads are variable and when customers are still learning which tasks should be automated at all.

This is not the first time the AI market has split between infrastructure economics and application economics. Cloud computing went through a similar phase, with customers initially paying for raw resources before demanding higher-level managed services and better cost controls. In AI, the distinction is sharper because the most visible products are conversational systems, while the enterprise purchasing decision often depends on far less visible features such as governance, auditability, latency management, and model routing.

That makes Karp’s comments strategically useful for Palantir. If enterprise buyers become skeptical of open-ended token spending, companies that package AI into higher-level systems may gain leverage. But the reverse is also true: if OpenAI and Anthropic continue improving model quality fast enough, many customers may accept token pricing as the cost of access to the best capabilities.

What we know, and what remains unclear

The strongest confirmed fact from the available source material is that CNBC reported Karp “bashes OpenAI, Anthropic token model” and quoted him saying “Something has gone completely wrong.” Without the full article text, there is not enough source evidence here to reliably state the full context, whether the comments were made during an interview, earnings-related appearance, conference discussion, or another public setting.

There is also not enough evidence in this source item alone to say whether Karp proposed a specific alternative pricing structure, cited customer examples, or directly referenced products from OpenAI or Anthropic. Those details would matter because “token model” can describe a billing mechanism, but Karp may also have been criticizing broader incentives around how AI companies measure value.

It is important to separate confirmed reporting from market interpretation. CNBC appears to have established that Karp made the criticism. The broader conclusion—that Palantir is using this moment to advance its enterprise AI positioning—is analysis based on Palantir’s known market strategy, not a direct quote from the source excerpt provided.

Similarly, any suggestion that token pricing is failing in the market would go beyond the evidence available here. OpenAI and Anthropic remain central suppliers in generative AI, and token-based billing is still widely used across APIs and developer tools. The news is not that the model has collapsed, but that a prominent enterprise software executive is openly arguing that it is misaligned with customer needs.

Implications for AI builders and enterprise buyers

For builders, Karp’s comments highlight a practical issue: product design increasingly determines cloud and model costs. Teams building on OpenAI or Anthropic can ship quickly, but they also inherit the economics of frequent inference calls, long context windows, and multi-step AI agents. In consumer products, that may be manageable if engagement is high and margins support it. In enterprise AI, especially internal productivity systems, the return on each token-heavy workflow may be harder to prove.

That creates pressure to optimize architecture. Builders may respond by using smaller models more often, reserving premium models for harder tasks, adding caching layers, narrowing prompts, or moving some workloads to open-weight alternatives. Others may look to orchestration vendors that promise better controls across model selection, retrieval, and policy management. In that sense, criticism of token pricing may accelerate demand for platforms that abstract model usage rather than expose every inference cost directly to the customer.

For enterprises, the message is even more concrete. Buying AI is no longer just a question of which model performs best in a demo. It is a procurement issue involving budget predictability, vendor dependence, security, and operational accountability. A team may like OpenAI for coding help, Anthropic for policy-sensitive use cases, and Palantir for deployment into internal systems. Those choices increasingly interact.

This is especially relevant for companies evaluating AI agents. Agents can multiply model calls behind the scenes because they plan, retrieve context, call tools, and verify outputs. A token-based pricing structure can therefore produce cost spikes that are not obvious during pilots. That does not make the model wrong, but it does make observability and workflow design more important.

Palantir’s critique may resonate most with buyers already frustrated by uncertain inference bills. Still, model vendors have their own defense: token pricing reflects actual usage and avoids forcing light users into expensive fixed contracts. For some organizations, that flexibility is a feature, not a flaw.

Evidence and claim quality

This story rests on a thin source base. The single available item is a CNBC report surfaced through Google News, and the extracted text does not include the full article body. As a result, the article can responsibly report that CNBC says Alex Karp criticized the token-based approach associated with OpenAI and Anthropic, and that he used the phrase “Something has gone completely wrong.”

Beyond that, caution is necessary. There is no direct evidence here of financial impact, customer reaction, or any new Palantir product move tied to the criticism. There are also no benchmark claims or adoption metrics in the provided source material. Any comparison of Palantir, OpenAI, and Anthropic in this article is therefore about business-model positioning rather than verified performance or market share.

What to watch next

The next signal to watch is whether Palantir turns Karp’s criticism into a more explicit commercial argument inside AIP and related enterprise AI sales messaging. If the company starts emphasizing cost governance and alternative pricing structures more aggressively, that would suggest this is not just commentary but a coordinated market push.

A second signal is how OpenAI and Anthropic adapt their enterprise packaging. Both companies already offer a mix of API and business-facing products, and sustained pressure from large customers could encourage more fixed-fee, seat-based, or hybrid pricing options alongside token billing.

Third, watch how AI agents are sold into enterprises over the next few quarters. If buyers become more sensitive to hidden inference costs, vendors that can show clear unit economics and usage controls may gain an advantage over those relying on broad automation promises.

Finally, it is worth tracking whether more enterprise software CEOs join Karp in criticizing pure usage-based pricing for large language models. If that happens, the debate could shift from a Palantir talking point into a broader market correction.

Creati.ai perspective

Karp’s comments matter because they expose a structural issue that the AI market has not fully resolved: the companies creating the most advanced models and the companies deploying AI into real institutions are often optimizing for different things. Model vendors want flexible monetization tied to compute-intensive usage. Enterprises want predictable costs, governance, and systems that solve business problems without turning every workflow into an open meter.

The likely outcome is not the end of token pricing. It is a more layered market. OpenAI and Anthropic will continue to sell model access where performance leads. But as enterprise AI matures, more value may move to orchestration, controls, and packaged workflows—areas where Palantir, and others building around enterprise AI deployment, believe they have an advantage. The real question is not whether tokens disappear. It is which vendors can make token consumption feel economically manageable inside production software.