American Firms Are Skipping the Best AI Model—and That Changes Everything

The Flagship Model Problem

Anthropic released Claude Opus 5 to widespread acclaim. By every benchmark, it was the best AI model on the market—a fact even competitors acknowledged. And yet, two months after launch, American enterprises largely ignored it.

Financial Times reported that Opus 5 accounted for just 11.4 percent of Anthropic’s total enterprise AI spending as of late July. Meanwhile, the older Opus 4.8 captured 34.9 percent, and the newer but cheaper Opus 5 took 19.1 percent. Companies that once line-danced behind every flagship release were quietly choosing not to.

The reason is bluntly economic. Opus 5 charges $10 per million input tokens and $50 per million output tokens. OpenAI’s GPT-5.6 Sonnet—performing noticeably below Opus 5 on head-to-head benchmarks—costs half as much: $5 per million input, $30 per million output. For a business running thousands of document summaries, data整理 tasks, and routine customer service queries, that difference is not academic. It is the gap between profit and loss.

The Dual-Strategy Becomes Standard

What is emerging is not a rejection of advanced AI but a recalibration of where it belongs. Enterprises are deploying a dual-strategy model: reserve flagship systems for genuinely complex work—legal analysis, scientific research, multi-step reasoning—and route everything else through cheaper mid-tier or older models.

An industry source told the Chosun Ilbo that summarizing documents or organizing data does not require cutting-edge capability. The metric that now matters, they said, is not peak performance but how cheaply a company can secure the performance it actually needs.

This shift has immediate consequences for the companies that have spent years winning mindshare by pushing benchmarks ever higher. Anthropic, OpenAI, and their peers built their enterprise narratives on capability leadership. Now the metric buyers apply is different—and it rewards a different kind of product strategy.

Opus 5’s launch did not move the needle because the needle had already moved. Mid-range models from both Anthropic and OpenAI improved fast enough that the performance ceiling became less urgent than the price floor. Opus 5 overtook Opus 5 in enterprise spending almost immediately after its release, a stark signal that the market is pricing itself out of premium models whether the providers like it or not.

The Price War Is Already Underway

The response from AI providers has been rapid and unequivocal. OpenAI cut the price of its lighter GPT-5.6 Luna model by 80 percent last month and reduced its Terra mid-tier by 20 percent. In August, it announced that GPT-5.6 Sonnet would see a further 20 to 33 percent discount through November 21. Anthropic reportedly scrapped plans to raise the price of its Sonnet 5 mid-tier model by 50 percent starting in September. DeepSeek offered its V4 Pro flagship at half price during weekends and off-peak weekday hours.

These are not promotional gestures. They are structural adjustments to a market that has decided performance alone no longer justifies the premium.

The pricing pressure is also changing product strategy. OpenAI’s Sonnet model—priced well below its flagship—now accounts for 25 percent of OpenAI’s total enterprise token usage. The cheapest option is not losing market share; it is capturing it.

Chinese Models Ride the Cost Curve

The most consequential side effect of this shift may not be between American providers but between American and Chinese ones. Chinese open-weight models—Kimi K3 from Moonshot AI and GLM-5.3 from Zhipu AI—have closed the performance gap to Anthropic and OpenAI’s top models while running at roughly one-third the price.

On OpenRouter, the token usage share of Chinese models exceeded 60 percent last month. That is a dramatic reversal from early 2025, when they captured less than 10 percent. The trajectory is clear: as enterprises optimize for cost-performance rather than absolute capability, Chinese open-source models gain distribution not through superiority but through accessibility.

Morgan Stanley noted in a recent podcast that companies are actively reducing spend on cutting-edge AI and shifting toward cheaper Chinese open-source alternatives. The firm framed this as a budgetary adjustment. But it is also a structural realignment of the global AI supply chain.

Who Wins, Who Loses

The winners in this environment are enterprises that can structure their AI deployments around tiered pricing—retaining flagship access for rare but critical tasks while running the bulk of operations on affordable models. They are also the Chinese open-source developers who are positioning themselves as the cost-effective backbone of global AI adoption.

The losers are the flagship model providers who priced themselves above the enterprise comfort zone. Anthropic’s Opus 5 is a technical achievement; commercially, it is struggling to find its buyer. OpenAI faces the same dynamic with its Sonnet lineup—strong but perpetually overshadowed by its own pricing architecture.

For the broader compute arms race, the signal is ambiguous. Lower token prices do not necessarily mean less compute consumption; they could mean more. If cheaper models unlock adoption at scale, total demand may rise even as per-unit revenue falls. But the revenue-per-token compression is real and ongoing, and it changes the calculus for every company betting on AI infrastructure.

What Comes Next

The enterprise AI market is entering a phase where benchmark leadership is necessary but no longer sufficient. The companies that survive this transition will be the ones that master tiered product strategy—offering a credible flagship, a competitive mid-tier, and an economical entry point—rather than those that treat performance optimization as the only lever worth pulling.

Chinese open-weight models will continue gaining share as long as the performance gap remains within acceptable bounds for most enterprise workloads. That window could narrow if American providers close the capability gap at equivalent prices, or it could widen if Chinese models continue their improvement trajectory.

What is certain is that the era of benchmark-chasing as a standalone enterprise value proposition is over. The next chapter will be written by whoever can deliver the right performance at the right price—not the highest performance at any price.