DeepSeek is becoming the darling of the AI landscape. This series of models, created by a Chinese research and development lab of the same name, has accomplished what once seemed impossible. It’s competing with models from OpenAI and Meta at a fraction of the cost. But is that really the case?
18 times cheaper than GPT-4. The Chinese startup launched DeepSeek V3 671B in late December 2024. South China Morning Post and the Financial Times report that this massive model was trained in just two months with a budget of $5.58 million. Its performance is comparable to OpenAI’s GPT-4, which CEO Sam Altman said it cost around $100 million to develop. As such, DeepSeek V3 is nearly 18 times less expensive to produce.
|
|
COST OF 1M INPUT TOKENS |
COST OF 1M OUTPUT TOKENS |
PRICE INCREASE |
|---|---|---|---|
|
CHATBOTS |
|
|
|
|
DeepSeek V3 |
$0.14 |
$0.28 |
|
|
Llama 3.1 405B Instruct |
$0.90 |
$0.90 |
542% / 221% |
|
OpenAI GPT-4o |
$2.50 |
$10 |
1,686% / 3,471% |
|
Gemini 1.5 Pro |
$1.25 |
$5 |
793% / 1,686% |
|
Claude 3.5 Sonnet |
$3 |
$15 |
|
|
GPT-4o mini |
$0.15 |
$0.6 |
7% / 114% |
|
MODELS WITH “REASONING” CAPABILITIES |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
COST OF 1M INPUT TOKENS |
COST OF 1M OUTPUT TOKENS |
PRICE INCREASE |
|---|---|---|---|
|
MODELS WITH “REASONING” CAPABILITIES |
|
|
|
|
DeepSeek R1 |
$0.55 |
$2.19 |
|
|
OpenAI o1 |
$15 |
$60 |
2,627% / 2,639% |
Incredibly cheap. The cost of DeepSeek’s API is remarkably lower than that of its competitors. When considering the prices of DeepSeek, Meta, OpenAI, Google, and Anthropic, it becomes evident that using DeepSeek through its API is significantly more affordable than the options available from its rivals. Although we’ve included the cost of GPT-4o mini in the tables above for comparison, its performance lags behind that of DeepSeek V3.
While DeepSeek V3 is superior to many of its competitors, it’s worth noting that Meta has recently launched Llama 3.3, and comparisons in the AI space are often subject to rapid changes.
(Theoretically) superior to all other models. According to a Reddit user, the pricing for DeepSeek V3 is promotional. Starting Feb. 8, costs will increase to $0.27 per million input tokens (almost double) and $1.10 per million output tokens (nearly four times more). This pricing change makes competitors look more appealing, especially Llama, which is the only rival that can match these costs. However, many benchmarks suggest that the Chinese model outperforms Meta’s and most others.
DeepSeek also “thinks” more cost-effectively. DeepSeep isn’t only cheaper when compared to traditional chatbots but also to “reasoning” models. According to DeepSeek’s internal benchmarks, the impressive DeepSeek R1 significantly surpasses OpenAI’s o1. However, using o1’s API costs 27 times more than DeepSeek R1–an astonishing difference.
A price drop is on the horizon. Expert Ethan Mollick predicts a price drop in the market as a result of DeepSeek’s influence. He estimates that the cost of an AI model like GPT-4 has decreased by 1,000 times in 18 months. He adds that a 95% reduction in the cost of reasoning models (currently more expensive than the AI models underlying ChatGPT) is quite feasible.
A Chinese revolution. The launch of DeepSeek’s models presents a significant opportunity for developers of AI-based solutions. They can now access models that are much cheaper and either equal to or superior to their competitors. This shift puts rival companies in a challenging position, and it’ll be interesting to see how they respond.
Great news for users. In the end, reduced prices make accessing AI tools much more affordable for both users and developers. The market has been moving in this direction, but DeepSeek has made a sudden and significant leap in cost reduction.
Image | Engin Akyurt
View 0 comments