Anthropic has officially changed the pricing structure for its Claude Enterprise plan, moving to a model that more closely mirrors cloud computing economics.
Under the new system, enterprise customers will pay a base fee of $20 per user per month, plus additional charges for every token they use at standard API rates. This marks a significant departure from the previous bundled tiers, which included a set amount of discounted tokens. In essence, the financial risk of high usage now shifts from Anthropic to its customers, a common practice in the cloud services industry.
This strategic shift is primarily driven by two key factors. First is the persistent compute scarcity in the market. The hardware that powers advanced AI, such as GPUs and high-bandwidth memory (HBM), is in short supply. This supply crunch makes running AI models—a process called inference—increasingly expensive. By billing for consumption, Anthropic can align its revenue directly with these fluctuating operational costs.
Second, the change reflects a broader trend of competitive normalization. Other major players like OpenAI are also segmenting their user base more granularly, introducing new pricing tiers to charge heavy users closer to their actual consumption. Anthropic's move completes this transition for its largest customers, ensuring its pricing remains competitive while being sustainable.
The economic impact on customers will vary widely based on their usage patterns. For light to moderate users, the new model could result in lower monthly bills. For example, a team using about 12 million input and 6 million output tokens might pay around $146, a notable saving from the old $200 bundled price. However, for 'power users', the costs could increase substantially. A heavy-usage scenario with 40 million input and 20 million output tokens could lead to a bill of approximately $440, more than double the previous fixed cost.
- API Token Rates: The price a customer pays for processing data with an AI model, typically measured per 1,000 'tokens' (pieces of words). Input (the user's prompt) and output (the AI's response) are often priced differently.
- Compute Scarcity: A situation where the demand for computing resources, like GPUs needed for AI, exceeds the available supply, leading to higher costs and limited access.
- HBM (High-Bandwidth Memory): A specialized type of memory that provides very fast data access, crucial for training and running large AI models efficiently.