DeepSeek-V4-Flash official version public beta: single task cost is about 60% lower than GPT-5.6 Luna

The official version of V4-Flash of DeepSeek started public beta on August 2. The selling point of this grade is very focused - maximizing the "comprehensive cost of a single task". The official comparison figures are straightforward enough: The cost of a single task is about 60% lower than that of GPT-5.6 Luna, and it claims a cache hit rate of about 98%, which reduces the reasoning overhead of repeated context scenarios such as multiple rounds of dialogue and similar Prompt batch calls.

The competitive anchor is moving from "capacity" to "unit cost"

V4-Flash continues DeepSeek's consistent line of "low cost + open source" and is positioned as a cost-effective level in the V4 series for cost-sensitive high-frequency calls. Looking at the 60% cost difference alone, it is easy to read it as a simple "price war"; but its real significance is to pull the focus of market discussion from parameters and scores to "comprehensive cost of ownership per unit task" - for large-scale callers, this often affects selection decisions more than a few tenths of a point on the list.

The 98% cache hit rate amplifies this logic: repeated context runs at almost zero marginal cost, which is especially critical for Agent multi-round orchestration and batch data processing workloads, and is almost a watershed for large-scale implementation. The two observation points worth keeping an eye on during the public beta phase are high concurrency stability and **consistency with complex tasks (long codes, deep reasoning)—the former determines whether it can handle real traffic, and the latter determines how difficult tasks it can handle.

It is worth mentioning that DeepSeek's recent actions do not stop at the model layer: from the open source of inference infrastructure to the self-construction of computing base, they are all progressing simultaneously. The public beta of V4-Flash is more like a concentrated display of the "cost leadership" flywheel on the inference side, and it has also pushed the price pressure on closed-source high-priced models to a new height.

Several directions worth tracking in the future:

1. High concurrency test during public beta period: Stability data under real load is a hard indicator of whether it can be used in the production environment.
2. Complex task consistency: The performance of long codes and deep reasoning scenarios determines the upper limit of its capabilities.
3. Commercial pricing and open source weight: Formal commercial pricing and subsequent open source rhythm will directly affect the developer selection pattern.