High AI Token Prices Are Pushing Users to Chinese Alternatives

John Authers used his Monday Bloomberg Opinion column to argue that the AI boom’s hottest chapter has run out of road. The piece, headlined “A Token Grasp of the AI Boom Shows Trouble Brewing,” runs with a photo caption that puts the thesis plainly: “High token prices could push users to lower-cost Chinese alternatives.” Authers framed the moment as the end of “the period of vertical ascent.”

Authers’s Column Pins the Boom’s Cost Problem

The column appeared on July 6 in Bloomberg’s Opinion newsletter under Authers’s byline (dateline: July 6, 2026 at 9:00 PM PDT). Bloomberg did not make the column body available for this report, but the headline, subtitle, and photo caption confirm the argument: the AI industry’s growth curve is bending, and the reason is cost, not capability. Authers’s piece on the AI token squeeze sits inside a stack of reporting that documents the same shift from other angles.

Western AI labs built their businesses on flat-fee subscriptions that charge a set monthly price regardless of how much computing a user consumes. Behind that flat fee, the marginal cost of serving heavy users has risen with usage. The gap between price and cost has narrowed from both directions over the past year. Users are responding to it.

The Subscription Math That Won’t Hold

SemiAnalysis, the research firm that tracks AI infrastructure economics, bought every Anthropic and OpenAI subscription plan on offer and stress-tested each one with long-running coding tasks until the weekly limit kicked in. The results, published in June, break the flat-fee model open. They are the basis for the cost complaint Authers’s column is making.

A user who fully maxes Claude Max 20x, Anthropic’s $200-a-month plan, would consume about $8,000 of tokens at API rates. The same exercise on ChatGPT Pro 20x, OpenAI’s $200 tier, lands around $14,000. Anthropic breaks even on its lower Claude Pro and Claude Max 5x plans at 20% utilization, per SemiAnalysis, and OpenAI starts losing money on its base plans once usage passes 11.4%. The top tiers are worse: Anthropic hits 0% gross margin at 10% utilization, and OpenAI slips into the red at 5.7%.

The firm also flagged a single unnamed company that burned through $500 million in one month after failing to cap employee license usage. Big spenders including Microsoft, Meta, and Amazon have started backing off “tokenmaxxing,” Tom’s Hardware reported, citing SemiAnalysis. WSJ reporting cited by Tom’s Hardware found that allowing agents to switch between AI models could cut costs by up to 95%.

Powerful agentic AI workloads consume up to a thousand times more tokens than the average query, per the same SemiAnalysis note. The subscription stress test that exposed the math laid out the rest.

How Chinese Labs Got 10x Cheaper

On the other side of the price gap, Chinese model developers are not inching down the cost curve. They are cutting the floor out from under it. Xiaomi, on May 27, rolled out cuts of up to 99% on its MiMo-V2.5 series and reset every active Token Plan so existing customers received five to eight times the credits they had paid for. The conglomerate, led by chief executive Lei Jun, has said it will invest at least $8.7 billion in AI by 2029.

DeepSeek made its month-old V4-Pro discount permanent at the same time. API prices for V4-Pro now sit in a range of 0.025 to 6 yuan per million tokens, roughly $0.0035 to $0.83, down from a prior range of 0.1 to 24 yuan. The lab, founded by hedge fund billionaire Liang Wenfeng, is also opening its cap table to external investors for the first time, with reports of a $3 to $4 billion round at a valuation of up to $50 billion led by the state-owned Chinese semiconductor fund “Big Fund III.” The 99% cut and permanent V4-Pro reset are documented in detail.

You don’t need a model that knows quantum gravity. These open-source models are very capable, and the ability to charge a big premium for AI is going to diminish.

The quote came from Vishal Misra, Columbia University’s vice dean, via the Wall Street Journal and reported by Tom’s Hardware. The pricing wedge lets Western buyers split the workload between cheap and frontier tiers. On OpenRouter, MiMo-V2.5-Pro lists at $0.435 per million input tokens and $0.87 per million output tokens. Flo Crivello, the founder of the AI assistant startup Lindy, told the Wall Street Journal he moved his company onto DeepSeek V4 after finding it “as capable as Sonnet while costing ten times less.”

Who Is Switching, and What They Save

Crivello is one of a small but growing list of US-based builders willing to be named. Stu Clott, an operations manager and part-time developer in San Diego, told the account of developer migration to Chinese AI models that he moved from Claude to DeepSeek and now spends less than 50 cents on an hour of coding that would have cost him about $10 on Claude. Ruben Garcia Jr., a Dallas-based developer who builds websites and mobile apps for businesses, told the outlet he pays $500 a month for Claude and ChatGPT to handle the hard planning and review work. Garcia spends another $200 a month for Minimax, Moonshot’s Kimi, and Xiaomi MiMo to handle the 90% of tasks that fall into coding, voice recognition, and routine processing.

The pattern is climbing the customer stack. Microsoft is exploring DeepSeek or another open-source model as a lower-cost alternative for Copilot Cowork, Rest of World reported. Copilot Cowork is currently powered by Anthropic and OpenAI systems. The frontier tier is kept for cases where capability gaps still matter, the cheap tier for everything else.

Adoption Has Outrun Revenue Conversion

The usage signal is already loud. Vercel, whose AI Gateway routes traffic from thousands of apps, said DeepSeek’s share of token usage on its network jumped from under 1% to 17% in May, while its share of revenue stayed near 1%.

The production index tracking token traffic to DeepSeek documented the shift. On Vercel’s index, 22.2% of AI Gateway requests ended with a tool call in April 2026, up from 11.4% in October 2025. Measured by tokens, the shift is bigger, at 58.9%.

On OpenRouter, a marketplace that lets developers route between models, the four most popular models are now all Chinese:

• DeepSeek
• Tencent
• Minimax
• Xiaomi

Kyle Chan, a fellow at the Brookings Institution, framed the gap for Rest of World: “AI adoption at large companies is fairly saturated. The growth market for Chinese companies would be medium-sized businesses that are starting to get into AI but are wary of the costs.” The conversion problem is structural. The same cheap tokens that buy usage do not yet buy the enterprise trust that turns a one-off integration into a multi-year contract.

The Political and Compliance Friction

The same dynamic that drives adoption also keeps revenue on the table. Lawmakers launched investigations into Airbnb and Anysphere, the owner of coding platform Cursor, after both companies disclosed using Chinese open models such as Qwen and Kimi in their AI infrastructure. Airbnb chief executive Brian Chesky later clarified that the company was not sending data to the model developers.

Lindy founder Flo Crivello told users on X that his company accessed DeepSeek through an American provider. He addressed a commenter who raised the question of whether user data was being sent to China. Some US firms run open-source models on their own servers. Others route through American cloud providers that may or may not share revenue with the Chinese labs.

Poe Zhao, the Beijing-based founder of the China tech newsletter Hello China Tech, told Rest of World that companies in highly regulated industries will stay reluctant to use Chinese models over data security, censorship, and geopolitical concerns. The cloud-routing arrangement prevents Chinese developers from building long-term enterprise relationships with American clients. Usage is only the first step, he said, and revenue conversion has yet to follow.

What the Western Labs Are Doing About It

OpenAI is now considering significant price cuts as it competes with Anthropic for enterprise users, the Wall Street Journal has reported. Chief executive Sam Altman acknowledged the problem in public remarks, saying the company was looking for ways to help users “get more value for less spend” when using ChatGPT. SemiAnalysis is forecasting that serving Opus 4.8-level models at $20 a month could become profitable as new models arrive and more data centers come online. Frontier offerings like Mythos would stay on per-token API pricing.

The trajectory Authers flagged now has a clock on it. Rest of World quoted Dallas developer Ruben Garcia Jr., who put the math plainly: “If the Chinese models come out and they are frontier and cheaper, I’m going that direction.” The boom is no longer being carried by capability alone, and Chinese labs have set the price.

Frequently Asked Questions

Why are AI token prices pushing users to Chinese models?

OpenAI and Anthropic charge a fixed monthly subscription, regardless of how many tokens a user consumes. The underlying cost behind that flat fee can be many times higher for heavy users. When SemiAnalysis tested every tier in June, fully using Anthropic’s Claude Max 20x would consume about $8,000 in API-priced tokens, while the same test on OpenAI’s ChatGPT Pro 20x would land around $14,000. Chinese developers have answered by cutting API rates to a small share of those levels and releasing open weights, which has made DeepSeek, Xiaomi MiMo, and Moonshot’s Kimi the default for routine, high-volume work.

How much cheaper are Chinese AI models than US alternatives?

Crivello, founder of AI assistant startup Lindy, told the Wall Street Journal that moving from Anthropic’s Sonnet to DeepSeek V4 cut costs tenfold with no perceived loss in output quality. The price gap shows up at the per-token level: on OpenRouter, Xiaomi’s MiMo-V2.5-Pro lists at $0.435 per million input tokens and $0.87 per million output tokens, while DeepSeek’s V4-Pro range extends as low as roughly $0.0035 per million tokens for some text inputs.

What is DeepSeek V4’s API price per million tokens?

DeepSeek reset its V4-Pro pricing on May 27, locking in API rates of 0.025 to 6 yuan per million tokens. That range works out to roughly $0.0035 to $0.83, depending on whether the workload is text input or text generation. The reset made permanent a discount campaign DeepSeek had launched the month before, and the lab is now raising outside capital at a valuation of up to $50 billion.

Who has publicly switched from Anthropic or OpenAI to Chinese models?

Lindy, the AI work assistant startup, told the WSJ it cut over to DeepSeek V4 from Anthropic models, with founder Flo Crivello saying the move saved the company millions. Microsoft is exploring DeepSeek as a lower-cost option for Copilot Cowork, which currently runs on Anthropic and OpenAI systems. Two named independent developers, Stu Clott in San Diego and Ruben Garcia Jr. in Dallas, told Rest of World they already split workloads between US frontier models and Chinese alternatives. Clott uses DeepSeek; Garcia uses Minimax, Moonshot’s Kimi, and Xiaomi MiMo for the bulk of his coding and voice work.

Why are OpenAI and Anthropic losing money on subscriptions?

Flat $200 subscriptions subsidize heavy usage heavily. Per SemiAnalysis’s June test, OpenAI’s base plans run into the red above 11.4% utilization, and Anthropic’s top plan lands at 0% gross margin at 10%. Agentic AI workloads consume up to a thousand times more tokens than an average query, which is why one unnamed company burned through $500 million in a single month.

Are US companies allowed to use Chinese AI models?

There is no blanket US prohibition on using Chinese AI models, but using them carries political risk. Lawmakers launched investigations into Airbnb and Anysphere, the parent of Cursor, after both disclosed deploying open-weight Chinese models including Qwen and Kimi. Airbnb CEO Brian Chesky later clarified that no customer data was being routed to the model developers. Some US firms sidestep the issue by running open weights on their own servers, while others access the models through American cloud providers that may or may not share revenue with the Chinese labs.