DeepSeek-V4-Flash scores 52 at 11 cents

DeepSeek-V4-Flash’s real advantage is economics: capable enough for serious agent workloads, cheap enough to use at scale. The benchmark result is compelling, but developers should validate reliability and latency on their own task mix before routing production traffic.

• –The Intelligence Index averages nine evaluations spanning agents, coding, general capability, and scientific reasoning, offering broader signal than a single benchmark.
• –Its 284B-parameter architecture with only 13B active at inference helps explain the cost-performance balance.
• –At $0.11 per completed task, Flash becomes especially attractive for high-volume coding agents, automation, and backend workflows.
• –Open weights expand deployment flexibility, but the model’s total size still makes self-hosting a serious infrastructure decision.
• –Cost-per-task remains workload-dependent: long contexts, retries, tool calls, and cache-hit rates can materially change the economics.