The AI industry has entered the "economic accounting stage" for the first time.

OpenAI and Anthropic may be heading into a price war. On the surface, it appears to be two leading AI companies competing for users, but deep down, it marks the first time the AI industry is truly entering the "economic accounting stage."

Over the past two years, the most discussed topic in the AI industry has been model capabilities: whose model is stronger, whose reasoning is better, whose coding ability is more stable, and whose multimodal outputs are more stunning. Capital, media, and users are all willing to pay a premium for the "strongest model." But when AI truly enters the enterprise realm, the situation begins to change. Companies no longer only ask "Is the model powerful?" but start to ask "Is the bill expensive?"; no longer just look at "Is AI useful?" but begin to question "How much result did these tokens actually exchange for?" This is the real point worth observing in this news.

According to Reuters citing The Wall Street Journal, OpenAI is considering significantly lowering AI service prices to compete for users with Anthropic; discussions are still ongoing, and Reuters also notes they could not independently verify the report. The news also mentions that corporate executives are feeling pressure from AI usage costs, and Sam Altman recently admitted that costs have become a "huge problem." I think this can be viewed in the context of Anthropic's rise in the enterprise sector, Claude Code's popularity, OpenAI's focus on promoting Codex, enterprises beginning to control agentic AI spending, and the pre-IPO business model stress test. My judgment is: this is not a simple token price cut war, but the beginning of the AI industry shifting from "smart display" to "value accounting."

This is not a traditional price war

According to the logic of traditional internet economics, if one company lowers prices, others follow suit, users benefit, the market expands, and costs are spread over scale. But large model services do not fully follow this logic. AI is not just software distribution with near-zero marginal costs; it involves real capital expenditures on computing power, storage, networking, electricity, cooling, and data centers.

Therefore, AI cannot be infinitely discounted. Prices can be optimized, per-token costs can decrease, model inference efficiency can improve, caching, batching, and model routing can reduce actual usage costs, but the underlying resource consumption will not vanish into thin air. The official OpenAI pricing page shows GPT-5.5 API prices at $5 per million input tokens and $30 per million output tokens; Anthropic's official info states Claude Opus 4.8 regular usage costs $5 per million input tokens and $25 per million output tokens. This indicates that top flagship models still maintain high-value task price anchors and have not entered a "bottomless low-price" phase.

Thus, what is called an AI price war is more likely a structural price reduction rather than a complete price collapse. Consumer, developer, low-tier models, and high-frequency entry scenarios may be the first to see price cuts; core enterprise tasks, complex reasoning, high-reliability code, and high compliance and security scenarios will still command premiums.

The real issue is not whether tokens are cheap, but whether each token can generate enough value.

From token maxxing to value maxxing

In the previous article "LiLi Observation · Tech Review," we discussed token maxxing: companies and employees constantly increasing token usage to prove they are "using AI." But now, the discussion of price wars further indicates that the AI industry must shift from token maxxing to value maxxing. Token maxxing focuses on "how much AI is used"; value maxxing focuses on "how much value each AI call actually creates." Behind these two terms are two completely different industry logics.

If a company only looks at token consumption, it can easily produce a seemingly prosperous AI usage curve: employees are using it, call volumes are rising, and bills are increasing. But this does not necessarily mean productivity is improving. Truly mature companies will next ask: do these tokens reduce rework? do they shorten delivery cycles? do they improve code quality? do they enhance customer experience? do they lower costs in sales, customer service, R&D, operations, and management?

This is the core issue once AI truly enters the enterprise. In the past, AI companies sold "intelligent capabilities"; now, enterprise clients want to buy "verifiable results." This is also the most important change behind the price war: AI competition is shifting from "who is smarter" to "who is more cost-effective," and further to "who can deliver results."

Why does OpenAI feel pressure?

OpenAI's pressure comes from two directions.

On one hand, is Anthropic's rapid rise in the enterprise sector, especially the spread of Claude Code among developer and enterprise engineering teams. Code scenarios are among the easiest for AI to form high-frequency, high-value, and sticky use cases because they can be directly embedded into R&D workflows, affecting delivery efficiency. The uploaded materials mention that Anthropic's programming tool Claude Code has driven revenue growth, prompting OpenAI to prioritize Codex. This indicates that the AI battlefield is no longer just chat interfaces but has entered real enterprise workflows.

On the other hand, is the re-evaluation of AI costs by clients. Early enterprise AI adoption was often experimental—buy first, use first, explore first. But as usage grows and bills increase, management will inevitably shift from "Is AI useful?" to "Is it worth it?" Reuters reports that enterprise executives are already expressing dissatisfaction with high AI costs.

This is not an AI boom ending, but AI entering operational budgets. Innovation budgets tell stories; operational budgets focus on returns. At this stage, companies will no longer only ask if the model is strong but will inquire about unit task costs, vendor replaceability, measurable results, cost sharing, and value assessment. My judgment is: if OpenAI considers lowering prices, it does not mean its business model is invalidated; it means AI is moving from "strategic pilot" to "operational expense." Once it becomes an operational cost, clients will start calculating AI more calmly.

The true change brought by price wars is industry rules

What this competition will truly change is not just the price of a particular model, but the evaluation system of the AI industry.

First, model companies cannot rely solely on "the strongest model" to tell stories; they must prove that "unit intelligence cost" continues to decline. Those who can accomplish more tasks with fewer tokens, lower latency, and more stable results will have an advantage.

Second, enterprise clients will become more proactive in model composition. In the past, companies might have directly purchased the strongest model; now they will differentiate task types: complex reasoning with flagship models, routine customer service with lightweight models, internal retrieval with local models, code review with specialized models. Future enterprise AI architectures are likely to involve multiple models, multiple vendors, and multi-layered cost management rather than relying on a single model.

Third, application companies will find opportunities. Falling model prices will compress some of the premium of foundational model providers but will amplify the value at the application layer. Because clients ultimately care not just about cheap tokens but about stable, deliverable, governable business results. Those who can embed model capabilities into real industry workflows will turn "token costs" into "business value."

Fourth, investors will change how they view AI companies. In the past, they focused on model leaderboards, user growth, and valuation stories; in the future, they will look at gross profit structures, inference costs, customer retention, task completion rates, workflow penetration, and unit task economics. Reuters reports that both OpenAI and Anthropic are moving toward IPO processes, with OpenAI reportedly having secretly filed IPO documents; this means the public markets will scrutinize their revenue quality, cost structures, and customer stickiness more directly.

My judgment is: AI price wars will not end the high-value narrative of AI but will end the phase of "big stories just because you can call models."

Implications for China's AI industry

For China's AI industry, there are three key lessons from this development.

First, Chinese large model companies cannot just compete on parameters, leaderboards, and launch events; they must enter cost-efficiency competition earlier. Chinese enterprise clients are more sensitive to prices and more pragmatic about ROI. If a model has decent capabilities but high reasoning costs, poor stability, or unclear delivery loops, it will be hard to truly integrate into core enterprise processes.

Second, China's AI opportunities do not lie in simply copying OpenAI or Anthropic but in deep industry scenarios and workflows. Manufacturing, supply chains, cross-border trade, financial risk control, government-enterprise services, park operations, customer sales, finance, taxation, and legal processes—these are complex workflows that Chinese companies already face. Those who can embed AI into these processes will have the chance to turn model capabilities into industry service capabilities.

Third, Chinese companies should establish their own AI cost governance systems as soon as possible. Don't just track how much AI is used; measure the input-output ratio for each task, process, and department. In the future, AI management should not only be about "which models were purchased" but also about "which tasks are suitable for AI, which models are most cost-effective, who reviews results, how costs are allocated, and how value is assessed."

Especially important: cheap prices do not equal actual low costs. Research from Stanford, Berkeley, and others on reasoning model costs shows that listed API prices do not always reflect actual inference costs; in some model comparisons, lower-priced models may consume more thinking tokens, leading to higher total costs. This is especially critical for Chinese companies: choosing an AI vendor should not only consider the per-million-token price but also the total cost of input, reasoning, output, review, and completion for each task.

This aligns with my consistent view: once AI truly enters enterprise, the core of competition is not who chats better but who completes tasks more effectively; not who burns more tokens but who can turn intelligence into lower-cost, higher-quality, verifiable results.

Conclusion: Cheap is not the end, value is the goal

If OpenAI and Anthropic truly engage in a price war, in the short term, it’s a battle for users; in the long term, it’s a test of business models. It will drive down AI usage costs and accelerate enterprise adoption; but it will also force the entire industry to answer a more serious question: Is AI creating value, or just generating bigger bills?

Therefore, the real significance of this price war is not whether tokens will be cheaper, but that the AI industry is finally moving from "smart display" to "value accounting."

The truly valuable AI companies in the future may not be those with the best conversational models, but those that can turn every unit of intelligence consumption into task results, business efficiency, and industry value.

The next stage of the AI industry is not about cheaper tokens, but about making each token more valuable.