Six months ago, OpenAI and Anthropic could credibly argue that price-sensitive enterprises had no serious alternative. That argument is collapsing.
CNBC reported on May 20 that both companies are expected to file for IPOs at valuations north of $800 billion, built on the premise that enterprises will keep paying a premium for frontier models because no real substitute exists. The data CNBC assembled makes the opposite case. Artificial Analysis ran each lab’s most capable model through 10 identical evaluations and totaled the bill. Anthropic’s Claude came in at $4,811. OpenAI at $3,357. DeepSeek, the Hangzhou-based Chinese lab whose model triggered a U.S. tech selloff in early 2025, at $1,071. Claude costs nearly nine times more than the cheapest Chinese alternative for the same work.
The market response to that gap is the “advisor model” pattern, which Databricks CEO Ali Ghodsi described to CNBC in concrete terms. An enterprise deploys a cheap open-source model as the default. When that model hits a task it cannot handle, it calls out to a frontier model from OpenAI or Anthropic. Ghodsi said this approach cuts costs sharply. The companies reaching for it are not startups running experiments. They are enterprises with AI budgets that crossed $100,000 per month. CloudZero survey data cited by CNBC showed 45% of companies spending at that level in 2025, up from 20% the prior year.
The open-weight pressure comes from several directions at once. DeepSeek and other Chinese labs including Moonshot and Zhipu have shipped models matching frontier capability on coding and agentic benchmarks in the past four months. Qwen, Alibaba’s open-weight series, and Mistral, the Paris-based open-weight lab, are pushing into the same cost-performance space from the Western side. On OpenRouter, a developer model marketplace, Chinese models went from roughly 1% of usage in 2024 to more than 60% in May 2026.
The counterargument, and it is a real one, is that regulated industries will not touch Chinese models regardless of price. Cohere CEO Aidan Gomez told CNBC his company’s revenue grew sixfold last year selling into banks, defense agencies, and similar buyers for whom security compliance overrides cost. But this is a narrow segment. The broader enterprise market, where security rules are less rigid, is a different story.
Here is the tension the IPO filings will have to resolve. Anthropic itself published a policy paper in May acknowledging that U.S. models are only “several months ahead” of Chinese ones and that Beijing is “winning in global adoption on cost.” At the same time, Anthropic reported $10.9 billion in revenue for the second quarter of 2026, a number that signals pricing power has not yet collapsed. Both things can be true. Revenue can be growing while the structural case for premium pricing weakens underneath it. Public market investors price the future, not the trailing quarter.
The position here is that the revenue figures will not settle this question. The advisor model pattern is a structural substitution, not a temporary experiment. Once an enterprise’s orchestration layer is routing by task complexity rather than by vendor default, the incumbent’s pricing leverage disappears at the margin. The labs that benefit most from this shift are the open-weight providers and the Chinese labs, neither of which will appear in OpenAI’s or Anthropic’s IPO filings as the threat they actually represent.
For enterprise buyers deciding on AI contract length now, the practical consequence is direct: avoid multi-year commitments to frontier model pricing without a cost-optimization clause. The advisor model pattern is available today, the open-weight alternatives are capable enough for most non-sensitive workloads, and the labs going public have a structural incentive to lock in contracts before their S-1 disclosures force this pricing conversation into the open.
Reported by CNBC on 2026-05-20.
