What changed
DeepSeek has made V4 Pro generally available and is charging two sets of API rates depending on time of day. The current pricing page lists peak windows of 01:00–04:00 UTC and 06:00–10:00 UTC; outside those windows, cached input, uncached input and output are each priced at half the peak rate. The production model is listed as DeepSeek-V4-Pro-0813 with a 1 million-token context window and 384,000-token maximum output.
Why it matters
Most model-price comparisons treat token rates as static. DeepSeek's schedule makes latency tolerance economically valuable: batch jobs, indexing, evaluations, report generation and other deferrable work can materially reduce token spend simply by moving outside peak windows. It also makes cache-hit rate and output volume more important to routing decisions because each component has a separate meter.
The same model has two daily price regimes
DeepSeek lists V4 Pro at peak rates of $0.044 per million cached input tokens, $1.32 per million uncached input tokens and $3.96 per million output tokens. Off-peak rates are $0.022, $0.66 and $1.98 respectively. The two peak windows total seven hours per day, leaving the remaining 17 hours at the lower rates. The pricing page says billing uses the prevailing rate when a request is processed.
Scheduling becomes a routing primitive
For latency-sensitive requests, time-of-day pricing may be irrelevant. For queues and background work, however, a scheduler can classify jobs by deadline and defer eligible work until off-peak periods. Builders comparing DeepSeek with other providers should therefore model effective cost by workload timing rather than comparing only a single headline token price.
Caching still changes the economics substantially
The cache-hit rate is far below uncached-input pricing in both periods, so systems that repeatedly send long shared prefixes, tool definitions or context can benefit from both caching and off-peak execution. Cost monitoring should keep cached input, uncached input and generated output separate rather than collapsing everything into one blended token rate.
Performance context still needs workload-level testing
Independent reporting places V4 Pro among stronger current reasoning and coding models, but benchmark summaries do not establish that it is the best choice for a particular agent, language, latency target or tool-use workload. The useful comparison is task success at a measured cost and latency under the exact API configuration a builder will run.