DeepSeek Plans Mid-July V4 Release With Peak-Hour Pricing

Time-of-day pricing is common in electricity and cloud spot markets but new to frontier LLM APIs, and DeepSeek applying it signals that inference demand, not training, is now the binding constraint for a major lab. For practitioners, this turns prompt scheduling into a cost lever: the same V4 call can cost one rate at 1 p.m. and half that at 8 p.m. Pipelines that are not latency-sensitive, such as nightly evaluation runs, bulk document processing, synthetic-data generation, and offline agent workflows, become candidates to move into off-peak windows, while user-facing traffic concentrated in business hours absorbs the higher rate.

DeepSeek said on Monday that the official release of DeepSeek V4 is scheduled for mid-July, building on the V4 preview the company shipped in late April. The company describes the official build as carrying a standard 1-million-token context window across the entire model lineup, with improvements in agent-based task execution, mathematical reasoning, and code generation. Alongside the model, DeepSeek introduced a peak and off-peak pricing plan that takes effect with the official release. Per the company, peak hours run 9:00 a.m. to 12:00 p.m. and 2:00 p.m. to 6:00 p.m. daily, during which API usage is billed at twice the off-peak rate; off-peak pricing is unchanged from current levels.

The structure is a demand-shaping mechanism, not a uniform hike. By holding off-peak rates flat and doubling them only in two daytime blocks, DeepSeek is nudging price-sensitive, deferrable load away from its busiest hours to protect availability and tail latency for everyone else. Multiple developers reported receiving advance pricing notices from DeepSeek, which corroborates that the change is being rolled out to existing API customers rather than reserved for new V4 tiers. Teams that previously chose DeepSeek primarily on headline price should re-run their cost models against a realistic hourly traffic profile, because effective cost now depends on when calls land, not just how many tokens they consume.

Open questions remain on the exact per-token figures for V4 Pro and V4 Flash at general availability, whether cache-hit discounts stack with off-peak rates, and how the peak windows map for callers outside China's time zone. If the model lands on its mid-July target and the demand-shaping works, expect other capacity-constrained providers to study time-based metering as an alternative to blunt rate limits or across-the-board price increases.