DeepSeek Launches V4 Pro and Sharply Raises API Prices
nashnova research
DeepSeek launched its V4 Pro model and hiked peak API output pricing by 350% to ¥27 per million tokens (~$4), signaling a pivot from rock-bottom pricing to monetizing agent infrastructure.
How big is the price hike — and is it actually expensive?
V4 Pro peak output jumps from ¥6 to ¥27 per million tokens (~$4) — a 350% increase.
Cached input pricing rises from ¥0.025 to ¥0.3 — an 1,100% jump.
This means → DeepSeek is no longer chasing "cheapest in the market." It is anchoring prices at "cheap but not free" — even after the hike, it remains far below U.S. frontier models.
In plain terms = the loss-leader phase is over; the user base is big enough to start charging real money.
Why raise prices now?
Just this May, DeepSeek announced a permanent 75% price cut, triggering follow-on cuts from ByteDance, Tencent, and others — a full-blown Chinese AI price war.
Three months later, the reversal signals the low-price strategy served its purpose: user scale and the developer ecosystem are already in place.
This reflects a broader shift in China's AI industry — from "burn cash for market share" to "whoever turns profitable first survives."
How good is V4 Pro's agent performance?
DeepSeek says V4 Pro is purpose-built for agent tasks — AI autonomously executing multi-step, complex operations. It matches or beats rivals on some agent benchmarks.
But the overall gap remains: its Artificial Analysis Intelligence Index score is 53, below Kimi K3's 60.
Alongside the model, DeepSeek open-sourced the Harness framework — an agent dev toolkit under MIT license — letting developers mix and match models, tools, and sandboxes.
This means → the playbook is "charge for the model, give away the framework" — the open-source toolkit pulls developers in; V4 Pro's agent capability is where the money is.
Where do fundraising and custom chips stand?
DeepSeek is seeking a new funding round at a reported valuation of ~$74 billion, having already closed ~$7.4 billion in its first external raise.
It is simultaneously hiring chip-design engineers to advance an in-house AI chip program.
In plain terms = raise prices to improve cash flow on one side, raise capital to stockpile resources on the other — the hike is not about running low on money, it is about proving the business model works to justify a higher valuation.
Can this pivot actually work?
From "lowest price online" to "350% hike" — the core bet is that developers are now locked into DeepSeek's ecosystem and willing to pay more for agent capability.
This reflects the key test for the next funding round: whether post-hike retention and revenue growth can support a $74 billion valuation.
If agent performance genuinely leads and developers stay, this is a rational harvest after winning the price war. If users defect to rivals, the hike becomes the move that drove them away.
市场有风险，内容仅供研究参考，不构成投资建议。