For the past year, DeepSeek’s main selling point was simple: it was dramatically cheaper than anything coming out of the US. That advantage is shrinking. According to Engadget, the Chinese AI company is raising its API prices by roughly four times in connection with the release of its new DeepSeek V4 Pro model, effective August 16.
The new pricing puts DeepSeek V4 Pro at $3.96 per million output tokens during peak hours, up from $0.87. Off-peak usage drops to $1.98, half the peak rate. DeepSeek says this peak and off-peak structure is designed to allocate compute resources more efficiently. The V4 Flash model follows the same pattern: $1.32 per million output tokens at peak hours (up from $0.28), and $0.66 off-peak. The gap between peak and off-peak is consistent across both tiers, which suggests this is a real capacity management mechanism, not just a pricing optic.
There’s some context worth keeping in mind here. DeepSeek’s current rates were originally a promotional discount meant to expire on May 31. The company extended them indefinitely that month, which at the time read as a strategic move to build market share. Extending then reversing that decision a few months later doesn’t look great, but the business logic is clear enough: promotional pricing built the user base, and now the company needs to move toward sustainable unit economics.
So how does the new pricing stack up? Still pretty well, actually. Kimi K3, the flagship model from Chinese rival Moonshot AI, costs $15 per million output tokens. OpenAI’s top-tier GPT-5.6 Sol sits at $30 per million tokens. DeepSeek V4 Pro at peak hours is $3.96 against those numbers, which is still a significant discount. The one place DeepSeek loses ground is against OpenAI’s budget option: GPT-5.6 Luna comes in at $1.20 per million tokens, which is cheaper than DeepSeek V4 Flash at peak hours. For cost-sensitive developers who don’t need maximum capability, that’s now a genuine trade-off worth evaluating.
The broader pattern here matters. DeepSeek arrived as a low-cost disruptor that forced the industry to take Chinese AI labs seriously. Its R1 model earlier this year rattled US tech stocks and triggered real conversations about compute efficiency. But running a competitive AI inference business at near-zero margins was never going to last. The move toward tiered, time-sensitive pricing looks a lot like what cloud infrastructure companies have been doing for years, and it signals that DeepSeek is operating more like a serious commercial platform than a subsidized experiment.
For developers and companies currently building on DeepSeek’s API, this is a good moment to re-evaluate. The cost advantage over OpenAI and Anthropic is still real, but it’s narrower than it was. Anyone who chose DeepSeek purely on price should run updated numbers. And anyone considering a switch from OpenAI’s cheaper tiers should check whether the capability difference justifies the cost difference. The answer may still be yes. But that math just got more complicated.
