Summary of Key Points
Two of the leading AI giants in Silicon Valley, OpenAI and Anthropic, which previously stood apart from price wars due to their strong model capabilities, suddenly began competing by lowering prices to attract customers this summer. Behind this price war are increasing costs for businesses using AI (as Agents integrated into workflows result in Token consumption becoming a direct production cost), narrowing performance gaps between models, the pressure from cheaper Chinese models, and decreased AI inference costs, all of which have created room for price cuts. While price wars save users money and provide more opportunities for experimentation, model companies face higher costs and financial pressures. In the future, only a few companies with substantial resources will be able to remain in the competition.
I. Silicon Valley AI Giants Competing on Price: From “Who is Smarter” to “Who is Cheaper”
Previously, both OpenAI and Anthropic viewed high prices as a sign of strong capabilities. For example, Anthropic’s Fable5 model cost $10 per million Tokens for input and $50 per million Tokens for output (with 1000 Tokens approximately equivalent to 750 English words or 500 Chinese characters), which was twice the cost of ordinary models. However, starting in July this year, the two companies turned against each other:
- Anthropic took the lead: They launched Opus5, cutting prices in half ($5 for input and $25 for output), claiming it had similar capabilities to Fable5—although its performance in programming tests was only 0.5% lower, while costs were reduced by half; even for complex tasks, the cost was just one-third of Fable5’s.
- OpenAI quickly followed: They released GPT5.6, with its flagship Sol model priced similarly to Opus5 ($5 for input and $30 for output), and specifically highlighted Sol’s cost-effectiveness by showing it used fewer Tokens to complete the same tasks.
- Packaging Plans Also Entered the Race: OpenAI reset usage quotas for every additional 1 million users (which could be saved for future use); Anthropic extended the benefits of its Fable5 packages despite increased computing demands. Both companies no longer solely emphasized model capabilities but focused on calculating “how much can be achieved with each dollar.”
II. Why Are Top Companies Suddenly Competing on Cost-Effectiveness?
There are four main reasons why leading companies are now engaging in price wars:
1. Businesses Can No Longer Afford High AI Costs: With Agents integrated into workflows, Token consumption has become a direct production cost. For example, advertising giant WPP uses more Agents than employees, and many AI-related expenses exceed their budgets; Uber burned through its annual AI budget in just four months. Companies now choose models based on who can complete tasks at the lowest cost.
2. Narrowing Performance Gaps: There are already many “good enough” models for everyday programming and data processing tasks. A16z’s survey found that 37% of companies use more than five models simultaneously, making model performance a comparable factor in choosing a solution.
3. Pressure from Chinese Models: Models like Kimi K3, which have recently gained popularity, are close to global leaders in coding and mathematical reasoning but cost half as much as Silicon Valley models ($3 for input and $15 for output), lowering user expectations for expensive models.
4. Decreased Inference Costs: Stanford data shows that the inference cost of GPT3.5-level models has decreased by 280 times over two years, and hardware costs have dropped by 30% annually. Improvements in model architecture and chip efficiency have created room for price cuts.
III. Users Benefit Significantly: Saving Money and Expanding Opportunities
The direct beneficiaries of these price wars are users:
- Real Savings: The difference in cost per million Tokens may seem small, but for businesses using billions of Tokens annually, the savings can be substantial. For startups and individual developers, this can determine the viability of their projects.
- Increased Experimentation: Higher costs previously prevented the use of advanced AI services (such as AI customer service or complex automation), but now these are more feasible. Companies can also mix models: using expensive ones for complex tasks and cheaper ones for routine work, allowing Agents to be tried multiple times until they perform correctly.
- Diverseer Products: Cheaper model “raw materials” lead to a wider range of applications. For example, while simple Q&A could only be done with ordinary models before, advanced models can now be used for complex data analysis and code generation.
IV. The Hard Times for Model Companies
Price cuts come at a cost to the companies:
- Soaring Costs: OpenAI’s inference costs have quadrupled in one year, reducing its gross margin from 40% to 33%. By 2030, its computing expenses are expected to reach $750 billion. Only by increasing usage faster than prices can companies cover their costs.
- A Capital Marathon: The price war has turned into a race to see who can sustain losses for the longest. If users stop using cheaper models or if margins are too low, increased usage will only exacerbate financial pressures. Small companies will not be able to survive; only those with significant capital and technology will remain.
- Shorter Periods of Premium Status: New capabilities in advanced models are quickly matched by competitors or integrated into cheaper models. For instance, Fable5’s advantage was short-lived as it was soon replaced by Opus5.
V. Future Trends: AI Will Become More Accessible, but the Bar to Entering the Field Will Rise
This price war will drive the AI industry in two directions:
1. Generalized Abilities Becoming Infrastructure: Basic AI capabilities will become as affordable as utilities, with only advanced models for complex tasks (such as long-text reasoning and advanced programming) maintaining a premium.
2. Shift in Competition Focus: The focus will shift from model capabilities to cost control, iteration speed, and user experience. Companies that can reduce inference costs more efficiently and better integrate AI into workflows will have a competitive advantage.
3. Narrowing of the Field: Only a few companies with capital and technology to withstand financial pressures will remain in the AI industry. For model companies, every day they stay in the game means spending more money.
In summary, this price war marks a transition for the AI industry from a focus on capabilities to an emphasis on efficiency. While users benefit, the competition for model companies will become more intense. Ultimately, AI will become more widespread, but the barriers to entering the field will increase.