Companies are shifting away from “tokenmaxxing” toward “valuemaxxing” as AI costs surge. Despite a significant drop in unit token costs over the years, the rise of agentic workflows has increased the number of tokens being used, driving up AI costs for many companies.
As a result, companies are recalibrating their AI budgets and looking into cheaper alternatives to frontier models, such as China’s DeepSeek. Meanwhile, third-party vendors have emerged to help companies alternate between different AI models, depending on their needs, which has helped them save money while still getting high-quality results.
In this article, we use findings from AlphaSense to examine the current state of AI token efficiency, how companies are responding, and what it means for AI vendors.
Related Reading: AI's Pragmatism Era: What It Means for Enterprises
AI Token Pricing
The unit cost of a single token has dropped significantly over the past three years due to intense competition between U.S. frontier labs and more cost-efficient alternatives, particularly those in China. Despite the drop in token prices, the numbers still differ significantly between the frontier models and Chinese alternatives.
For example, OpenAI’s flagship model GPT-5.5 charges $30 per one million output tokens. China’s DeepSeek model is considerably cheaper at 87 cents per one million output tokens. The drop in prices is good for customers. However, the advancements in AI models have resulted in more tokens needed for each task.
As AI models have advanced from simple chatbots to complex, multi-step agentic workflows, the number of tokens consumed has increased by up to 30x. As these AI systems have become more complex, providers have begun moving away from a flat subscription rate to per-token metered billing. In April, Anthropic transitioned its enterprise tier to a usage-based, per-token billing model.
I actually think the market is going to bifurcate. Commodity AI workloads like summarization, classification, retrieval, they're going to get cheaper. Premium frontier reasoning models will still come in a premium... From an enterprise perspective, we're planning on lower unit costs over time and higher overall consumption as AI becomes embedded in more workflows.
Even as token prices go down, companies are concerned about the amount of money they are spending on AI due to the introduction of agentic AI systems and how many tokens those systems use.
Companies Pull Back Token Usage
Uber, Accenture, and Walmart are just a handful of the companies who have recently pulled back their AI usage due to skyrocketing invoices. Uber burned through its entire 2026 AI budget just four months into the calendar year and has since put a $1,500 monthly limit on its employees token spending per AI tool. Accenture reportedly has been restricting its employees’ token use by preventing them from using it for basic tasks like converting PDFs into presentation slides. And Walmart is limiting staff’s usage of an in-house AI tool.
These actions show how companies are grappling to rein in costs after incorporating AI into their workflows. At the same time, some industry experts say that companies will soon begin to understand how much token usage their employees need to improve workflows and that the finance departments will be better able to predict their budgets moving forward.
I think clients, once they understand it, they start to come around because they realize they won't have to pay for what they aren't using, so once they get to know their consumption rates, they can really have some good cost controls and be more effective with their spending.
AI routing platforms are quickly emerging as a key player in solving the “tokenmaxxing” crisis, allowing users to dynamically allocate workloads across models based on cost and performance. These platforms can analyze queries and automatically direct them to the most suitable model, whether that be a high-cost frontier model like GPT-5.5 or a cheaper model like DeepSeek V4. Companies are increasingly adopting this strategy because it allows them to toggle between different models, providing them with a better way to manage their AI budgets.
Cheaper Competition
The competition between U.S. frontier labs and Chinese challengers like DeepSeek is heating up. While U.S. labs continue to lead in absolute frontier capabilities, experts argue that Chinese labs are demonstrating an ability to match near state-of-the-art performance at a fraction of the reported development cost.
The challenge from Chinese labs has triggered a reassessment of global AI capex and token pricing models globally. Despite spending only 18% of what American hyperscalers have invested, Chinese models are now benchmarking broadly in line with U.S. peers, with the lag on performance narrowing to one month, according to broker research. Even though the Chinese models are just shy of having the AI capabilities of the frontier models, analysts say they’re “good enough” for most tasks like coding and data processing.
If you're looking for the best open-source, you've got to go to China. DeepSeek R1 and V3.2 are cost-effective, and they are super strong with reasoning and math and open weights.
While some U.S. companies are adopting the Chinese models due to their lower costs, others are holding off due to security and governance concerns.
From a numbers perspective, it makes sense, but that's not necessarily what the full cost is. I think, first and foremost, you'd have to look at your data jurisdiction. Companies handling whether it's U.S., E.U., or sensitive user data, they can't risk that regulatory blowback. That's where even with fine-tuning on Chinese cloud infrastructure, if they're not local, that actually becomes a serious issue, which leads into where I would go with security and trust.
Still, notable companies like Airbnb, Cursor, and Coinbase use Chinese models, which has prompted U.S. lawmakers to look into how to curb the rise in adoption. In April, the House Committee on Homeland Security and the House Select Committee on China said they will jointly investigate the growing adoption of Chinese-developed AI models, beginning with letters to Cursor and Airbnb.
Balancing Token Costs With AI Value
The AI landscape is shifting so rapidly that companies are being forced to adapt in real-time. New models are emerging alongside a shift in pricing, as vendors move away from flat subscriptions toward consumption-based models tied to token usage. At the same time, Chinese labs are pressuring U.S. frontier providers to justify their price premium. Together, these changes mean companies need to remain flexible with their AI strategies so they can ensure they’re using the best technology at the right price.
Move from research to execution around the clock with AlphaSense's SuperAnalyst, an always-on AI agent purpose-built for financial and strategic workflows. Sign up for early access or start your free 2-week trial of AlphaSense today.
