Claude and GPT Cut Prices for Reusing the Same Material

Anthropic and OpenAI both released new models in September, and both cut the price of feeding the same material back into a model. Claude Opus 5.5, released on September 22, charges 60% less than its predecessor to re-read stored input. GPT-6.1 Sol, which arrived on September 29, halved the price of the same item compared with the previous version. Earlier in the month, on September 10, DeepSeek launched V4.1-Flash as a model developers can plug into their own services, releasing the model weights and a technical report at the same time.

What does it cost to reuse the same material?

Caching is the practice of storing content a model has already processed so it can be reused later. For any service that repeatedly feeds in long documents, this line item drives the total bill. Claude Opus 5.5 charges $0.20 per million tokens to re-read stored input. A token is the unit of text used for billing. Ordinary input runs $4 and output $20, and the model is also available on Claude’s paid subscription plans.

GPT-6.1 Sol’s standard pricing is $2 per million tokens for ordinary input and $10 for output. Stored input costs $0.10, down from $0.20 for GPT-6 Sol. Writing content into the cache costs $2.50. Once input exceeds 272,000 tokens, a higher long-context rate applies to the entire request. Users working with long documents need to watch length as well as price.

DeepSeek’s official price list shows V4.1-Flash charging $0.006 per million tokens for stored input during peak hours. Input that has not been stored costs $0.30 and output $1.20. During off-peak hours, each rate is cut in half. The structure means pricing depends not only on whether content is cached but also on when you use it.

What exactly did DeepSeek shrink?

DeepSeek says it reduced the memory footprint of the intermediate calculations a model holds while generating an answer. According to the company, it combined techniques that let multiple computation steps share and reuse the information they have stored. DeepSeek reports storage of 890 bytes per token, which it describes as roughly a quarter of what the earlier V4-Flash required.

V4.1-Flash accepts text and images together and can read up to one million tokens at a time. DeepSeek announced that it is ending support for V4-Flash and an experimental image version, and that existing API names will point to the new model for a transition period. The released model carries an MIT license. Running it yourself, however, requires substantial hardware to match the model’s size.

An opinion piece dated September 29, “The AI Race Just Got Awkward,” connected DeepSeek’s technical disclosure to the price changes at its two rivals. The piece argued that Anthropic and OpenAI copied DeepSeek’s caching techniques. Neither company explains its internal caching architecture in its official price list.

Is it available in South Korea?

South Korea is among the countries officially supported by both Anthropic and OpenAI. Claude lets users select Korean in the language settings on web and desktop, and OpenAI documents multilingual support in its models. DeepSeek’s V4.1-Flash model can be downloaded and run locally. Korean-language interface support and the quality of a model’s Korean-language answers remain separate questions.

These price changes feed directly into cost calculations for AI services that reuse the same material over and over. On top of ordinary input and output rates, users now have to weigh the conditions attached to stored input, long-context surcharges, and the time of day they run their jobs. How often the stored-input rate actually applies in real workloads, and how good the Korean output is, are the things worth watching next. Whether these cuts for developers eventually show up in consumer subscription prices is another open question.