DeepSeek Quietly Routes V4 Pro Traffic to Cheaper Flash Model, Users Push Back

DeepSeek appears to have started serving at least some traffic addressed to its V4 Pro endpoint with V4.1 Flash instead, according to developer accounts collected on the Chinese forum OSCHINA. The reports describe the substitution as unannounced, and the thread quickly filled with engineers comparing what their calls were actually returning. DeepSeek has not published a statement explaining the routing.

The two models occupy different rungs of the company's lineup. V4 Pro has been sold as the higher-capability option, while V4.1 Flash gives up some depth in exchange for lower latency and a smaller bill per token. Sending Pro-bound calls to Flash would trim DeepSeek's serving costs and take pressure off the larger model, which matters as inference demand climbs. It also means a request tied to one set of weights can come back answered by another.

The frustration voiced on OSCHINA centers on disclosure rather than quality. Several developers wrote that if Flash is cheaper and quicker, DeepSeek could have said so and let customers switch on their own schedule or A/B the two. What they object to is discovering the change by inspecting outputs. One recurring complaint: it is impossible to reproduce a result, hold a fixed benchmark, or explain to an end user why quality moved between two identical calls.

For teams shipping products on a named checkpoint, the model identifier functions as a contract. Quietly altering what sits behind that string breaks the contract even when the replacement costs less. The issue is sharper for anyone who tuned prompts or evaluation suites against Pro's specific behavior, since those artifacts no longer describe the system that answers them. Support tickets and regression tests lose their reference point.

The episode arrives while Chinese labs are competing aggressively on price per token, and DeepSeek has been among the most willing to cut. Smaller and distilled models have become a common way to absorb high-volume traffic without paying flagship rates on every call. That math is sound, but it runs into how API customers expect versioning to work. Alibaba's Qwen and Moonshot have drawn comparable questions about model swaps, generally with more warning beforehand.

What happens next hinges on whether DeepSeek documents the routing or rolls it back. Developers want to know when a Pro request will be handled by Flash, whether that can be turned off, and how pricing and output stability are affected. Until those answers arrive, anyone building on V4 Pro is working without a guaranteed baseline, and some teams are already running side-by-side tests against competing APIs.