OpenAI and Anthropic slash prices on same day, intensifying AI model price war

Two Industry Leaders Make Unprecedented Moves

On September 22, within roughly 90 minutes of each other, OpenAI and Anthropic—two companies that arguably need customers the least—announced significant price reductions for their latest models.

OpenAI unveiled GPT-6 Sol and GPT-6 Luna, while Anthropic launched Opus 5.5. Both companies prominently featured "affordability" in their marketing materials.

• GPT-6 Sol: Input price reduced to $2 per million tokens, output to $10 per million tokens—half the cost of its predecessor.
• GPT-6 Luna: Input price slashed to $0.10 per million tokens, with reductions of up to 58%.
• Opus 5.5: Input priced at $4 per million tokens, output at $20 per million tokens.

This coordinated price-cutting by a trillion-dollar valuation AI company with the largest user base, alongside the model provider with the highest enterprise adoption rate in the U.S., raises a critical question: What is driving this shift?

Two Divergent Pricing Strategies

Despite acting on the same day, OpenAI and Anthropic are pursuing fundamentally different approaches.

OpenAI: Tiered Pricing for Volume

OpenAI's strategy is "tiered pricing for scale." GPT-6 Sol targets complex coding and agent workflows, while Luna is designed for cost-sensitive, high-volume tasks.

The logic is clear: assign different model tiers to different jobs. Astra's flagship handles high-difficulty tasks, Sol manages routine coding and batch processing, and Luna processes high-frequency, lightweight conversations.

Crucially, OpenAI has stated these are permanent prices, not promotional or launch discounts. This signals a bet that sufficiently low prices will drive exponential growth in daily call volumes, with economies of scale ultimately making unit costs sustainable.

Anthropic: Total Cost Optimization

Anthropic's approach is markedly different. While Opus 5.5's per-unit price drop is only 20%, the company emphasizes "total cost of ownership." A new dynamic cache read mechanism reduces cache read prices from $0.50 to $0.20 per million tokens—a 60% decrease.

Additionally, Opus 5.5 requires roughly half the total tokens to complete tasks, combined with over 30% faster output. This results in approximately 40% lower overall operating costs for typical workloads compared to the previous generation.

To use an analogy: OpenAI is like a supermarket slashing shelf prices in half to attract customers. Anthropic is like a restaurant optimizing its supply chain—menu prices barely change, but each dish uses fewer ingredients and arrives faster, giving customers more value for the same money. One cuts unit price; the other cuts the total bill.

Three Structural Forces Driving the Price War

Why now? While competition is the surface-level answer, three deeper structural forces are at play.

Force 1: Breakthroughs in Technical Architecture

These price cuts are not subsidized losses. Anthropic's Opus 5.5 introduces a dynamic cache read mechanism that reduces intermediate state data storage and retrieval costs to one-quarter of previous levels. Combined with this underlying improvement, enterprises using the same core business data consume roughly half the tokens.

OpenAI's aggressive pricing on the Sol series similarly rests on a fundamental restructuring of computational logic. Luna further strips away non-essential general-purpose computing modules, focusing on high-frequency, short-turnaround conversational tasks.

When the underlying computing paradigm shifts, lower prices become a natural byproduct of new architecture—not a reluctant concession.

Force 2: Market Data Reveals a Harsh Reality—Expensive Models Don't Sell

Data from enterprise spending platform Ramp shows that Anthropic's most expensive Fable series accounted for only 6% of enterprise token purchases from Anthropic in its first month.

This simple statistic reveals a fundamental truth: most workloads don't require top-tier models. If a task can be completed with Opus, enterprises have no reason to pay more for Fable.

Meanwhile, Ramp data shows the effective price per million tokens has fallen from $1.15 in March to $0.68—a 41% decline in six months. Enterprises are becoming increasingly cost-conscious, no longer blindly paying for the "best model." Instead, they ask: "Can a cheaper model do this job?"

Force 3: Chinese Models Apply Downward Price Pressure

Chinese models, led by DeepSeek, have pulled down the global pricing floor for large language models by an order of magnitude. DeepSeek V4 Flash charges just 1 RMB per million tokens for input and 2 RMB for output—approximately $0.14 and $0.28, respectively.

Even with OpenAI's aggressive Luna price cut, its output price remains over four times that of DeepSeek V4 Flash.

Research indicates that while average API prices for Chinese large models have risen notably over the past year, U.S. closed-source model prices have been declining. The price gap between Chinese and American models is narrowing.

This "you rise, I fall" divergence suggests U.S. companies feel the pressure. If prices don't come down, price-sensitive customers will not hesitate to switch to Chinese models.

The Price War Across the Pacific: Different Playbooks

Looking back at China, this script played out two years ago.

In May 2024, ByteDance's Doubao led the charge, pricing its Pro model at 0.0008 RMB per thousand tokens—99.3% below the industry average. Alibaba Cloud's Tongyi Qianwen followed with a 97% price cut on its flagship model, Baidu's ERNIE Bot made two major models completely free, and Tencent's Hunyuan model saw price reductions of up to 87.5%. The industry descended into a frenzy of "selling tokens at a loss."

But two years later, the trend has reversed. Securities firm estimates show China's total daily token consumption surged from approximately 100 billion in early 2024 to 180 trillion by February 2026.

Amid this demand explosion, AI companies including iFlytek and Tencent Cloud have announced price increases, with some products rising over 400%.

By Q2 2026, the average API input price for Chinese large models had risen to 4.9 RMB per million tokens, with output at 21.9 RMB—increases of approximately 48% and 80% respectively compared to Q1 2025.

The price war has become unsustainable. A ByteDance executive reportedly stated internally: "In the next 18 months, only players who control the computing supply chain will survive."

This assessment applies to the U.S. market as well, but the underlying logic of the price wars in both countries differs fundamentally:

• China's price war is "burning cash for market share." ByteDance leverages pre-stocked computing advantages and industry-lowest marginal costs to crush competitors. Alibaba and Baidu are forced to follow—Alibaba, for instance, uses MoE architecture to compress 397 billion parameters into 17 billion active ones, reducing memory usage by 60% and increasing inference throughput 19-fold. This is "cost reduction under duress." The problem: when markets shift from growth to consolidation, burning cash becomes waste, not investment.

• The U.S. price war resembles "technology dividend distribution." OpenAI and Anthropic's price cuts are backed by clear architectural optimizations. They are not buying market share at a loss but recalculating costs under a new technological paradigm.

A recent report notes that Chinese AI companies are abandoning simplistic price wars, increasingly pricing based on task completion costs. In a sense, pricing strategies in China and the U.S. are converging toward the middle.

What Ultimately Determines the Winner?

Technical Scale

The foundation for price cuts is not fundraising capacity but the engineering capability required to reduce inference costs by even one percentage point.

Anthropic's cache read optimization and OpenAI's inference efficiency restructuring are technology moats built with real investment. Xia Lixue, CEO of Infinigence AI, notes that through coordinated optimization across models, system software, and hardware, inference costs have dropped 90% over the past year, with potential for another 90% reduction.

This means the endgame is not about who can withstand losses longest, but who can reduce costs fastest through technology. The faster cost reducer can maintain margins at lower prices; the slower one is forced out by losses.

Valuation

OpenAI is currently valued at $852 billion, while Anthropic's latest valuation reaches $965 billion—both approaching the trillion-dollar mark. These valuations enable sustained, astronomical investment in computing infrastructure.

ByteDance's 2025 capital expenditure exceeded 150 billion RMB, with approximately 90 billion directed to AI computing. Its 2026 plan allocates 160 billion RMB, including 85 billion for AI chip procurement.

Only at this scale can companies simultaneously cut prices and increase R&D investment. Small players simply cannot compete in this game.

Market Consolidation

The price war itself accelerates industry differentiation. Global startup funding totals approximately $510 billion, of which OpenAI and Anthropic alone have absorbed about $217 billion—43% of global venture capital.

With capital so heavily concentrated at the top, mid-sized and smaller model companies lack the technical capability to reduce costs, the capital to burn, and the confidence to sustain computing investment. When the price war ends, not all players will remain at the table.

Final Thoughts

In 2013, the subsidy war between Didi and Kuaidi burned through billions and ended in a merger.

In 2020, the community group-buying price war lasted over a year, leaving only two or three survivors.

The large model price war will be no different. It is an elimination contest about "who can deliver the best capability at the lowest price."

OpenAI and Anthropic's latest price cuts appear to be a price war, but they are fundamentally about using technological advantage and capital scale to redefine the industry's cost baseline.

As the price per million tokens falls below $1, large models are shedding their luxury status and becoming infrastructure as essential as electricity and water.

Only those with sufficiently solid technology, massive scale, and valuations capable of sustaining long-term investment will continue to operate and innovate at this price level.

The endgame of a price war is never about who is cheaper. It is about who can thrive while being cheap.

Technical scale and valuation are the threshold for survival.