The Rules of the Global Large Model Arena Are Being Rewritten

Image Source | Internet (Please contact for removal if infringing) Partially AI-generated

As the label of 'price destroyer' shifts its focus, the rules of the global large model arena are being rewritten.

Over the past two years, when the term 'price destroyer' for large models was mentioned, everyone thought in one direction: Chinese manufacturers engaged in cutthroat competition, driving API prices from the ceiling to the floor, and then even further to the basement.

On the other side of the ocean, Silicon Valley giants sat firmly at the other end of the pricing spectrum, guarding their premium pricing and enjoying a peaceful existence.

But this narrative was completely shattered in 2026.

Silicon Valley Giants Take the Initiative to 'Flip the Table'

Anthropic's newly released Claude Haiku 5.5 offers requests under 100,000 tokens at just $0.10 per million tokens for input and $0.50 for output, a tenth of the previous generation's price and even cheaper than the domestic small model DeepSeek-V4.1-Flash.

Almost simultaneously, OpenAI slashed the API price of GPT-5.6 Luna by 80%, reducing input from $1 to $0.20 per million tokens and output from $6 to $1.20.

This combination of moves was interpreted by experts from Southern University of Science and Technology as 'clearing the market with low-end volume, maintaining profitability in the mid-range, and locking in customers with high-end premiums.'

This essentially uses economies of scale to erode the survival space of smaller manufacturers, pushing general-purpose foundation models toward a 'utilities-style' affordable pricing model.

But if you've been following the industry closely, you might have noticed a detail: this is OpenAI's second price cut in the past month.

After adjusting prices for mid-to-low-end models late last month, they immediately followed up by 'slashing' their flagship series. Why would a company Sprint IPO ( Sprint IPO: preparing for an IPO) with a target valuation of $852 billion voluntarily compress its profit margins at this time?

According to a previous report by The Wall Street Journal, leading U.S. AI companies have even started giving away computing power for free to compete for enterprise users, to the point where some startups no longer need financing to purchase computing resources.

Does this narrative of giving away computing power for free sound familiar?

In fact, domestic manufacturers in China had already started using this tactic two years ago.

In other words, the current narrative of Silicon Valley giants regarding large model pricing is following the same path previously taken by domestic large model manufacturers.

But the deeper reason for Silicon Valley giants to 'lower their status' lies in the fact that the rise of open-source models is changing the rules of the game.

Domestic models like DeepSeek have driven the price of one million tokens down to a few jiao (Chinese currency unit) per million tokens and adopted an open-source strategy, allowing developers to deploy privately for free. This has eliminated the premium pricing space for API services.

In February 2026, the total token consumption of the world's top 10 AI models exceeded 27 trillion, with Chinese large models contributing 14 trillion, accounting for more than half.

This means that Chinese models are not just 'cheap' but are genuinely being used on a global scale.

When enterprise users find they can achieve similar results for a tenth of the cost, the 'performance halo' that Silicon Valley giants relied on to maintain high pricing is rapidly fading.

Forbes provided a judgment in an analytical article: the foundation model market is increasingly resembling the large-scale infrastructure industry, where scale, capital strength, and operational efficiency have become key to determining success. The core of the competition has shifted from 'how smart the model is' to 'how efficient the operation is.'

In plain terms, large models are transforming from 'luxury goods' to 'daily necessities.'

And the iron law of the daily necessities market is: whoever can reduce costs to the lowest will survive the longest.

From 'Burning Money for Volume' to 'Compensating Volume with Price'

Shifting our focus back to China. If Silicon Valley's price war is 'just beginning,' then China's price war has already completed a full round and even entered the 'second half.'

In May 2024, ByteDance's Doubao took the first shot, pricing its Pro model 99.3% lower than the industry average.

Subsequently, seven companies including Alibaba, Baidu, Tencent, and iFLYTEK followed suit, with 21 models successively reducing prices.

The keyword of that stage was 'burning money for volume,' prioritizing developer adoption and market share above all else. Insiders revealed that in 2024, a leading manufacturer internally set an aggressive goal of not considering profitability for three years, with product pricing even below computing power costs.

But by 2026, the tone had suddenly changed.

First, DeepSeek announced in May a permanent price reduction for its V4-Pro model API to a quarter of the original price, followed by Xiaomi's MiMo series with a maximum reduction of 99%, and Tencent Cloud's intelligent agent development platform offering up to a 97.5% discount for DeepSeek-V4 series models.

According to public information, in the first half of 2026 alone, the domestic large model industry witnessed six rounds of price cuts involving almost all major players including DeepSeek, Zhipu, Alibaba, ByteDance, Tencent, and iFLYTEK.

However, there is a fundamental difference from Silicon Valley's current round of price cuts: domestic manufacturers' price reductions have shifted from 'strategic losses' to 'technical cost reductions.'

Take Zhipu as an example. Its GLM-5.2 adopts a Mixture of Experts (MoE) architecture with 744 billion total parameters but activates only about 40 billion per inference. Combined with sparse attention mechanisms, it reduces memory usage by 75%, doubles throughput per machine, and halves the computing power cost per token.

The company's cloud business gross margin rose from 3.3% in 2024 to 18.9% in 2025—a figure more convincing than any price cut announcement: the price reductions are not about burning money but reflect genuine cost reductions.

Meanwhile, the pricing landscape of domestic large models is shifting from 'universal floor pricing' to 'tiered differentiation.'

Input prices for economy models have generally dropped to around 1 yuan per million tokens, but the price gap between flagship models is widening. For example, Kimi K3 charges as high as 20 yuan/100 yuan per million tokens for input/output, while MiniMax M3 costs just 2.1 yuan/8.4 yuan after a 'permanent 50% discount'—a cost difference of nearly 60 times for the same task.

This indicates that the domestic large model price war is no longer simply about 'who is cheaper' but has entered a stage of differentiated pricing and tiered competition.

Two Price Wars, Two Strategies

While Silicon Valley giants are intensifying price cuts, some 'reverse operations' have emerged in China.

Since the first quarter of 2026, Zhipu's API pricing has increased by 83% month-over-month, Tencent Cloud's Hunyuan model has seen price adjustments exceeding 400% for some services, and Baidu Intelligent Cloud's computing power products have simultaneously increased by 5%–50%.

Morgan Stanley's statistics are even more direct: in the second quarter of 2026, the average API input price for Chinese large models rose to 4.9 yuan per million tokens, with output prices reaching 21.9 yuan—increases of approximately 48% and 80%, respectively, compared to the first quarter of 2025.

With prices falling on one side and rising on the other, this set of data reveals a fascinating industry turning point.

After more than two years of price wars in China's large model market, the diminishing marginal returns of low pricing are becoming apparent.

Industry analysts point out that while the 2024–2025 price wars accelerated the market penetration of AI large models, they also left the industry mired in a 'high-investment, low-return' dilemma. When model invocation volumes jumped from hundreds of billions to trillions, computing power costs ballooned exponentially, making reliance on capital infusion unsustainable.

In other words, domestic manufacturers have discovered that being cheap alone cannot retain users.

When the capability gap between flagship models narrows to the point where users can barely perceive it, what truly determines developer retention is migration costs and ecosystem stickiness.

'The competition has entered a new phase, with the industry's competitive logic quietly shifting from 'price competition' to 'ecosystem competition,' ' an industry analyst told the media.

Leading manufacturers have already taken action: Zhipu is building an ecosystem combination of 'foundation models + programming tools + subscription services,' while Moonshot AI has tailored open-source licensing terms for Kimi K3 to shift its competitive moat from 'model weights' to 'hosting efficiency.'

When comparing domestic and international price wars within the same coordinate system, the differences become very clear.

Silicon Valley's current round of price cuts is essentially 'catching up.' Over the past two years, Chinese manufacturers have driven prices to rock bottom, while Silicon Valley's closed-source models relied on performance advantages to maintain premiums.

But as open-source models catch up in capabilities and the cost-performance balance tilts, Silicon Valley giants have had to lower their stance and engage in competition.

Anthropic's Haiku 5.5 not only slashed prices to a tenth of the previous generation but also addressed shortcomings in programming and computer operations. In the small model arena, 'being cheaper' is upgrading to 'being cheaper and better.'

China's price war, meanwhile, has moved past the 'brute subsidy' phase into the 'refined operation' deep waters.

Peak-valley differentiated pricing, cache mechanism optimization, MoE architecture cost reduction, subscription model exploration...

The underlying logic of these moves is the same: upgrading from 'selling capabilities' to 'selling services.' DeepSeek reduced cache hit prices during off-peak hours from 0.05 yuan to 0.02 yuan per million tokens—a reduction of over 60%.

This pricing strategy clearly guides developers to migrate batch tasks to off-peak hours, trading time for cost and diluting unit computing power consumption.

From a market development perspective, China's price war has evolved beyond 'who is cheaper' to 'who can profit while being cheap.'

This is the true watershed moment. From 'not considering profitability for three years' in 2024 to 'must consider sustainable operations' in 2026, the industry's mindset has undergone a fundamental shift.

After the Price War, What Comes Next?

Gartner predicts that by 2030, the inference costs of trillion-parameter large models for generative AI providers will decrease by over 90% compared to 2025.

AWS has also publicly stated that AI costs can drop by up to 100 times every few months.

As costs continue to decline, the price war will not end.

But in the end, victory will not hinge on who has the lowest price, as there is no bottom to price reductions.

The two variables truly worth watching are: first, who can maintain reasonable profit margins at low prices? Second, who can truly 'lock in' developers to their ecosystem?

Anthropic's approach may offer a clue.

Haiku 5.5 not only offers extremely low prices but also introduces a prompt caching mechanism, with reading prices inside the cache as low as $0.01 per million tokens. Combined with batch processing, this can save an additional 50% in costs.

The essence of this mechanism is to use technical means to cultivate user habits, making you unconsciously form path dependency while enjoying low costs.

In plain terms, the large model business ultimately competes not on technical parameters or price tags but on 'switching costs.'

Whoever can maximize switching costs will have the last laugh.