DeepSeek released the V4 Pro official version late on August 12, pushing Agent (AI that autonomously completes multi-step tasks) benchmarks close to—and in some areas past—top international rival Fable 5. API (the paid interface developers call) pricing remains at roughly one-tenth of its competitors. This marks the first time a Chinese large-model company has meaningfully closed the gap with the international first tier on the Agent dimension.
The model supports 1 million tokens (roughly 700,000 Chinese characters) of context input and 384,000 tokens of output, covering Thinking, Tool Calls, and the Responses API. Third-party testing shows V4 Pro can now autonomously complete three representative task types: analyzing 7,000 e-commerce data points to generate an interactive Dashboard webpage, turning the Odyssey into an interactive storytelling site, and building from scratch a universe sandbox with gravitational physics simulation—without meaningful human intervention across intermediate steps.
What this is
DeepSeek V4 Pro is the official flagship of the V4 series; its predecessor V4 Flash launched in late July. Core upgrades are concentrated in Agent capabilities: across Agent benchmarks including Terminal Bench and DeepSWE, V4 Pro approaches top rival Fable 5 on multiple metrics and surpasses it on two.
On pricing, official documentation shows a shift to peak/off-peak pricing starting August 17: 4.5 yuan/million tokens for input and 13.5 yuan/million tokens for output during off-peak hours, doubling during peak. Even at peak rates, input costs come in at roughly 1/8 those of Fable 5 and output at about 1/13.
One hidden shortcoming: V4 Pro lacks multimodal (understanding and generating text, images, video, and other content types) capabilities, so the visual design of generated webpages still requires an external image model to fill the gap. It can "do the work," but cannot independently "make it look good."
Industry view
Positive sentiment centers on value for money. Multiple tooling partners report that V4 Pro has, across data analysis, content understanding, and coding tasks, become capable of "filling in its own intermediate steps and carrying through to a finished result"—a meaningful cost-cutting tool for SMBs.
We find the reservations equally worth recording. First, the model is far from "set-and-forget"—visual and interaction details still need repeated human adjustment. Second, DeepSeek's pricing is in fact rising this round: a uniform rate has shifted to peak/off-peak, doubling at peak, so heavy users may see higher costs. Third, the third-party benchmarks come from platforms promoting their own integrated products, so assessment neutrality deserves a discount. Fourth, matching Agent leaderboard scores does not mean matching real-world task experience; leaderboard and production-environment performance often diverge.
Impact on regular people
For enterprise IT: The cost-effectiveness window is opening. Agent tasks that once cost over ten yuan per run now cost mere cents, dramatically lowering the cost barrier for outsourcing data analysis and reporting workflows to AI.
For working professionals: Non-coders can now let AI run multi-step tasks end-to-end, but a gap remains between "doing it itself" and "delivering a shippable product"—aesthetic and interaction judgments still need a human in the loop.
For the consumer market: The price war will eventually transmit to feature upgrades in consumer-facing AI tools, but personal willingness-to-pay is still in a cultivation phase—most users are still "kicking the tires" with free tiers.