On August 21, OpenAI announced a three-month price reduction for GPT-5.6 Sol: input tokens fall from $5 to $4 per million, and output tokens fall from $30 to $20 per million. Google cut Gemini 3.7 Flash pricing to half of its predecessor. But the same week, Chinese AI companies are moving in the opposite direction—DeepSeek's new peak-hour pricing went into effect August 17, with output tokens reaching 27 yuan per million, 4.5x the previous rate.
The divergence is not coincidence. It's a signal.
OpenAI's Discount Lasts Three Months. That's Not a Coincidence.
OpenAI's strategy is defensive. The company needs to keep developers locked into its ecosystem while it prepares for whatever comes next—likely an IPO, a major model release, or both.
The Sol price cut is temporary—three months, not permanent. That's a key distinction. OpenAI is not resetting its pricing floor. It's running a promotion. The company is betting that developers who switch during the discount will stay when prices return to normal—because the switching cost, once you've integrated deeply, is higher than the cost of the API.
Google is playing a similar game, cutting Gemini 3.7 Flash pricing to half. But Google's move is more defensive: Gemini's flagship Pro model is still missing, and Flash is the workhorse that keeps developers in the ecosystem. The price cut is a retention tool.
DeepSeek's Price Hike Is Permanent. That's the Point.
DeepSeek's move is not defensive. It's offensive.
The company raised peak-hour prices to 27 yuan per million output tokens—4.5x the previous rate. It also introduced peak/off-peak pricing, a model that treats compute like electricity: expensive during business hours, cheaper at night. The message is not "we're getting more expensive." The message is "our compute is in demand, and we're pricing accordingly."
DeepSeek is not alone. Zhipu, Kimi, and MiniMax have all raised API prices this year. The "price war" phase of Chinese AI is ending, and the "value capture" phase is beginning. As one analyst put it: "This isn't a unified industry-wide price increase. It's independent model vendors leading the charge, while large cloud providers hold steady."
Different Business Models, Different Price Moves
The difference comes down to business models.
OpenAI and Google are cloud-platform companies. Their AI APIs are loss leaders or margin-reduced products designed to pull developers into their broader cloud ecosystems. Google can subsidize Gemini pricing with GCP revenue. Microsoft can subsidize OpenAI with Azure. They can afford to cut prices to keep developers in the tent.
DeepSeek, Zhipu, and Kimi are pure-play AI model providers. They don't have a cloud business to cross-subsidize their model APIs. When compute costs rise, they have to pass them on. Their pricing has to reflect real costs—and real demand.
The deeper signal is that Chinese AI companies believe their models are now good enough to command higher prices—and that developers will pay. That's a shift from "we're cheaper, try us" to "we're competitive, stay with us."
What It Means
A year ago, every AI company was cutting prices. Now the market is splitting. U.S. companies are cutting temporary discounts to defend share. Chinese companies are raising permanent prices to capture value.
Neither strategy is wrong. They reflect different stages of maturity, different business models, and different competitive pressures. But the divergence itself is the signal: the AI pricing war is no longer a single front. It's two separate battles—and both sides think they're winning.
P.S. OpenAI's price cut lasts three months. Google's Gemini discount runs to 2027. DeepSeek's peak pricing is permanent. The time horizon tells you everything about the confidence each company has in its position.
Frequently Asked Questions
Q: What did OpenAI do to its API pricing?
A: On August 21, OpenAI announced a three-month price reduction for GPT-5.6 Sol: input tokens dropped from $5 to $4 per million (20% cut), and output tokens dropped from $30 to $20 per million (33.3% cut). The discount runs for three months. ChatGPT Work and Codex credit packages also got price cuts, but Pro, Plus, and Business subscriptions were unaffected.
Q: What did DeepSeek do to its API pricing?
A: DeepSeek's peak-hour pricing went into effect August 17. Peak output pricing reached 27 yuan per million tokens—4.5x the previous rate. The company also introduced peak/off-peak pricing, treating compute like electricity: expensive during business hours, cheaper at night. DeepSeek is not alone—Zhipu, Kimi, and MiniMax have all raised API prices this year.
Q: Why are OpenAI and DeepSeek moving in opposite directions?
A: OpenAI and Google are cloud-platform companies. Their AI APIs are loss leaders designed to pull developers into their broader cloud ecosystems. They can subsidize AI pricing with cloud revenue. DeepSeek, Zhipu, and Kimi are pure-play AI model providers. They don't have a cloud business to cross-subsidize their APIs. When compute costs rise, they have to pass them on.
Q: Is OpenAI's price cut permanent?
A: No. The GPT-5.6 Sol price cut runs for three months. OpenAI is running a promotion to attract developers, betting that the switching cost will keep them locked in after prices return to normal.
Q: What about Google's Gemini pricing?
A: Google cut Gemini 3.7 Flash pricing to half of its predecessor. The discount runs to 2027. Google's move is defensive—its flagship Pro model is still missing, so Flash is the workhorse that keeps developers in the ecosystem.
Q: What does the divergence signal?
A: A year ago, every AI company was cutting prices. Now the market is splitting. U.S. companies are cutting temporary discounts to defend share. Chinese companies are raising permanent prices to capture value. The divergence signals that the AI pricing war is no longer a single front—it's two separate battles.
Q: Which companies are raising prices in China?
A: DeepSeek, Zhipu (three times this year), Kimi, and MiniMax have all raised API prices. As one analyst noted, "This isn't a unified industry-wide price increase. It's independent model vendors leading the charge, while large cloud providers hold steady."
Q: Why is DeepSeek raising prices while OpenAI is cutting them?
A: DeepSeek believes its models are now good enough to command higher prices and that developers will pay. The company is signaling that its compute is in demand. OpenAI, by contrast, is trying to protect market share while preparing for a major model release or IPO.
Q: What is the broader market signal?
A: The split reflects different stages of maturity and different business models. U.S. companies are playing defense with temporary discounts. Chinese companies are playing offense with permanent price increases. Neither strategy is wrong—they reflect different competitive pressures.
Q: What does the time horizon tell us?
A: OpenAI's discount lasts three months. Google's Gemini discount runs to 2027. DeepSeek's peak pricing is permanent. The time horizon reveals confidence: Chinese AI companies believe they can maintain their pricing. U.S. companies are buying time with discounts.
