DeepSeek Offers 75% Off-Peak Discount to Manage Demand Spike

 Hangzhou-based Chinese artificial intelligence (AI) star DeepSeek has announced a new discount programme, slashing prices for accessing its models through its application programming interface (API) during off-peak hours, in response to high demand that has strained its server resources during the day.

DeepSeek revealed the new rates on Wednesday, and they will take effect on Thursday after midnight.

From the hours of 12.30am to 8.30am China time, API access to the V3 model will be available at a 50 per cent discount. This comes to US$0.035 per million tokens for cache hits, US$0.135 per million tokens for cache misses, and US$0.55 per million tokens for output.

Access to the start-up’s R1 reasoning model will be available during the same hours at a 75 per cent discount. The two models are now priced identically during off-peak hours.

Context length – the maximum number of tokens a model can process at one time – is 64,000 tokens for both models. A token in AI refers to a fundamental unit of data processed by the algorithm, which can be a word, number, or even a punctuation mark. DeepSeek bills users based on the total number of input and output tokens processed by its models.

By comparison, OpenAI’s o1 reasoning model is priced at US$15 per million input tokens, US$7.5 per million cached input tokens, and US$60 per million output tokens, with a context length of 200,000 tokens.

The pricing adjustments come as DeepSeek continues to grow in popularity, leading to persistent server issues. Earlier in February, the company temporarily stopped developers from topping up their accounts, blaming “server resource constraints”. Existing balances could still be used.

On February 10, DeepSeek also ended a promotional campaign for accessing its V3 model via API. Meanwhile, the company’s official namesake chatbot has struggled with persistent server overload, disrupting continuous conversations for users.

After more than two weeks, DeepSeek resumed account top-ups on Tuesday and followed up with today’s announcement for off-peak discounts.

DeepSeek is currently in the middle of what it is calling Open Source Week, where it is releasing a new code repository for each workday. The company promised last week to make five of its repositories related to its AI work fully open source.

The company has released the code for FlashMLA, DeepEP and DeepGEMM, all aimed at maximising chip performance for cost-effective model training and inference tasks.

Reuters reported this week that DeepSeek is accelerating the release of its R2 model, which was originally scheduled for May.

https://www.scmp.com/tech/tech-trends/article/3300264/ai-night-chinas-deepseek-offers-peak-75-discount-demand-strains-servers