Large Model Price Cuts: OpenAI and Anthropic See Their Highest Price Reduction of 80% in Half Years

Opening Dialogue

Shiwen: In the past six months, OpenAI and Anthropic have successively reduced prices, with some models seeing price cuts of up to 80%. For ordinary developers, this may directly affect costs more than the release of new models. What do you think about this wave of price cuts?

Yongliang: My first reaction is that this isn’t just a simple promotion; it’s the entire industry re-pricing. In the past, the charging model for large models was somewhat like selling vegetables by weight—the more tokens used, the more expensive it became. Now, companies are starting to lower the unit cost, indicating an increase in upstream supply and growing competition.

Q1: Why did leading manufacturers suddenly reduce prices intensively starting from June?

Shiwen: Looking at the timeline, in June, OpenAI released GPT-5.6 Luna with prices of $1/6/0.10 per token; on July 30, they announced a 80% price cut to $0.20/1.20/0.02 per token. On August 21, they temporarily reduced the price of GPT-5.6 Sol by about 20%, to $4/20/0.40 per token. On September 22, they released GPT-6 Sol and Luna with prices of $2/10/0.20 per token and $0.10/0.50/0.01 per token respectively. On the same day, Anthropic released Claude Opus 5.5 at $4/20/0.20 per token, with the total cost reduced by 40% to 60% compared to Opus 5. Anthropic also released Fable 5.1 on September 1 at $10/50/0.25 per token. Why have there been so many concentrated price cuts in these months?

Yongliang: The core reason is that the market has entered a new phase. In the first half of the year, the gap in model capabilities was narrowing, and users began to vote on prices. Both OpenAI and Anthropic know that if they don’t follow suit in pricing, developers will easily switch to cheaper alternatives. Price cuts are both a defense measure and a way to capture the ecosystem. Additionally, on September 29, GPT-6.1 Sol halved the price of cache read to $0.10 per token. The cumulative price cut for the entire Sol series has reached 75%, indicating that these price cuts are not one-time events but a continuous signal.

Q2: Why is cache read price cutting more critical than input/output?

Shiwen: Many price cuts focus on cache read. For example, the cache read of GPT-5.6 Luna was reduced from $0.10 to $0.02 per token, and the cache read of Fable 5.1 was reduced from $1 to $0.25 per token. Why is everyone cutting on this aspect?

Yongliang: Because in long-context scenarios like agentic coding, cache read can account for 96% to 99% of total token costs. No matter how low the unit price of input and output is, it cannot withstand the high-frequency overhead of repeated context reading. Cutting cache read means directly reducing the most sensitive hidden cost for developers. For applications that handle a large number of historical sessions and code libraries daily, this savings is very significant. You can think of it as the toll on highways: the entrance fee and fuel cost remain the same, but the toll stations for repeated passages have been reduced, which is felt most by those traveling long distances.

Q3: What role have Chinese manufacturers and open-source heavyweight models played?

Shiwen: Isn’t the pressure to reduce prices just due to competition between OpenAI and Anthropic?

Yongliang: Of course not. The pricing for DeepSeek Flash V4.1 is $0.15/$0.60/$0.003, and Moonshot Kimi K3.1 Pro is even cheaper than its predecessors, at $2/$10/$0.20. These models have pushed the market price anchor down. With the deployability and customizability of open-weight models, leading closed-source manufacturers must prove their premium prices are justified; otherwise, users will choose open-source solutions. A new balance is forming between the low prices of Chinese models and the brand premium of Western models.

Q4: What does model commercialization and low switching cost mean?

Shiwen: When prices converge and capability gaps narrow, will large models become standard products like cloud servers?

Yongliang: That’s the trend. The switching cost between models is getting lower, and API interfaces are becoming more standardized, with abstraction layers maturing. For many companies, switching to a different model may only involve changing a few configurations. This means brand premium will decline, and what truly matters are the toolchains built around models, security governance, and industry solutions. In the future, the models themselves may be just a pipeline, with value added at both ends. Manufacturers need to retain customers not just by name but through end-to-end engineering capabilities and compliance guarantees.

Q5: How should developers and enterprises respond to this round of price cuts?

Shiwen: For those actually developing products, price cuts are beneficial, but there’s also a lot of noise. What advice would you give?

Yongliang: First, recalculate total cost of ownership. Don’t focus only on unit prices; cache read and hit strategies may determine the actual bill. Second, assess whether the most expensive model is truly necessary; many scenarios can be handled with mid-range models. Third, maintain architectural flexibility so your profit margins aren’t tied to a single manufacturer’s pricing. Fourth, seize the opportunities brought by model commercialization, such as multi-model routing and dynamic degradation. Price cuts will continue, and only those who can convert cost advantages into product capabilities will have long-lasting success.

Host: Shiwen (AI) | Guest: Yongliang