DeepSeek Just Quadrupled Its API Prices While Everyone Else Is Cutting Theirs

$0.28 per million output tokens. That was the price that made DeepSeek the default choice for developers who wanted frontier-quality AI without the bill. Starting 16 August, that number becomes $1.32 at peak hours. A 4x increase.

DeepSeek announced pricing changes across its API products this week, citing the need to "allocate resources more reasonably." V4-Flash output tokens jump from $0.28 to $1.32 per million during peak hours. V4-Pro goes from $0.87 to $3.96 per million. The company is introducing peak and off-peak pricing tiers for the first time, acknowledging that demand has outgrown its inference infrastructure.

The cheap AI narrative just got complicated

DeepSeek proved you could build frontier models on a fraction of the budget. Its V3 model, trained for reportedly $5.6 million, matched models that cost hundreds of millions to develop. That story made it the darling of the AI cost-efficiency crowd and drove a wave of adoption, particularly across Asia.

But training a model cheaply and running it cheaply at scale are two different problems. Inference costs scale with demand, and DeepSeek's demand has been growing faster than its capacity can match. The result is what any economist would predict: prices go up.

What makes this timing unusual is the direction everyone else is moving. OpenAI has been slashing prices throughout 2026. Anthropic dropped its Claude API costs twice this year. Google's Gemini pricing has become increasingly aggressive. The major Western labs are racing to the bottom. DeepSeek, which was already at the bottom, is now racing back up.

What the numbers actually look like

Even after the price hikes, DeepSeek remains cheaper than GPT-5.6 and Claude for most use cases. The gap has narrowed from roughly 10x cheaper to about 3x cheaper, which is still significant but no longer the obvious choice it once was.

For startups and developers who built their products around DeepSeek's ultra-low pricing, a 4x increase changes unit economics overnight. Some will absorb the cost. Others will start evaluating alternatives they previously dismissed as too expensive.

The peak and off-peak tier structure adds complexity. Developers now need to think about when their applications are calling the API, not just how often. That is a scheduling problem that did not exist when the price was uniformly low.

What this means for the AI pricing war

The broader pattern is revealing. Building frontier AI models has gotten cheaper. Running them at scale has not. Inference is the new bottleneck, and the companies that control GPU capacity are gaining leverage.

DeepSeek's IPO ambitions valued at $9 billion through its partnership with Unitree Robotics suggest the company sees its future in more than just cheap API access. But today's price increase is a reminder that the economics of AI infrastructure are still being figured out in real time.

Samsung's $648 billion AI chip investment and the broader semiconductor supply chain buildout are partly responses to this exact problem. Until new fab capacity comes online, inference costs will remain a constraint, and companies like DeepSeek will keep adjusting prices to match reality.

The era of "AI is getting cheaper every month" is not over. But it is more nuanced than the headlines suggested. Training costs drop. Inference costs depend on demand. And demand, right now, is winning.