OpenAI and Anthropic slash LLM prices by up to 80% to counter competition

A rapid race to the bottom

Over the last few months, the landscape for large language model (LLM) pricing has shifted dramatically. OpenAI and Anthropic have implemented sharp price reductions, with some models seeing costs drop by as much as 80%. These cuts appear to be a direct response to increasing pressure from open-weight models and a growing number of inference providers.

The timeline shows a clear pattern of tit-for-tat pricing adjustments between the two giants. For instance, OpenAI slashed the price of GPT-5.6 Luna by 80% in July, while Anthropic later made its temporary Sonnet 5 discounts permanent. These moves suggest that the leaders are struggling to maintain a price premium as the technology matures.

As minimaxir.com reported, these price cuts are not accompanied by a decrease in output quality. This indicates that the companies are finding ways to optimize their infrastructure rather than cutting corners on the models themselves. The goal is to maintain market share while the underlying technology becomes more accessible to developers.

The threat of open weights

A major driver of this price war is the rise of open-weight models, which provide sufficient performance for many tasks at a fraction of the cost. Models like DeepSeek Flash V4.1 have gone viral because they offer extremely low cache read costs. Cache reads are a major part of the total expenditure when using agents for coding, as they contain the history of the entire conversation.

Because open-weight models can be hosted by many different providers, the barrier to entry for the inference industry is low. Developers can easily switch to the lowest-priced provider via platforms like OpenRouter. This creates a commodity market where there is very little 'moat' or unique advantage for the major labs beyond their highest-performing models.

To counter this, OpenAI and Anthropic are focusing on higher-quality models like Opus 5.5, which requires less compute to serve. By making these top-tier models more efficient, they can offer lower prices while still differentiating themselves from the cheaper, open-weight alternatives. They are essentially trying to out-innovate the competition on both performance and price.

Infrastructure and efficiency

The price cuts also reflect genuine improvements in how these models are served. Anthropic explicitly stated that Opus 5.5 requires less compute to serve than its predecessor, and its pricing reflects that reality. This suggests that as the technology evolves, the cost to run a single inference becomes cheaper for the providers.

This efficiency is a double-edged sword for the industry. While it allows for lower prices for the end user, it could theoretically reduce the need for massive data center buildouts. However, because inference is only one part of the total compute mix, these data centers are still necessary for the training and development phases of the AI pipeline.

For the consumer, these efficiencies mean that high-end capabilities are becoming more affordable. The trend suggests that as long as providers can price tokens below the revenue they generate, they will continue to lower costs to stay competitive. This puts downward pressure on the 'true cost' of a token, potentially making it closer to the marginal cost of production.

IPO goals and profit margins

The timing of these price cuts is particularly interesting given the upcoming IPO plans for both companies. Anthropic is expected to go public within a few months, while OpenAI is targeting 2027. Usually, companies want to show high revenue growth before an IPO, but these price cuts might be intended to capture more volume and market dominance.

While the companies are losing billions due to R&D and scaling, reports suggest that the profit margins on inference itself remain healthy. By lowering the price of the primary revenue source, they may be trying to build a larger user base and establish a dominant ecosystem before the public markets take notice. They are moving toward a model where they own the platform and the primary tools used by developers.

To protect their margins, these companies are also focusing on enterprise commitments and personal assistants like OpenAI Dots and Claude Duet. These products help create a 'moat' that is harder to replicate than a simple API. By shifting the focus toward integrated services, they can maintain profitability even as the base price of raw tokens continues to crater.