DeepSeek just changed the equation: V4 API costs now depend on the clock
DeepSeek’s new V4 rates went live on August 16: peak windows cost twice as much as off-peak periods, while even the off-peak prices are substantially above the previous flat rates.
What changed
The new rates took effect at 16:00 UTC on August 16. DeepSeek now charges twice the off-peak rate between 01:00–04:00 and 06:00–10:00 UTC. V4 Flash output, previously $0.28 per million tokens, is now $0.66 off-peak and $1.32 at peak. V4 Pro output moved from $0.87 to $1.98 or $3.96.
Agent economics changed most
The more important change for agent systems is cached context. One million cached V4 Pro input tokens previously cost $0.003625; they now cost $0.022 off-peak and $0.044 at peak. That directly changes the economics of long sessions, repository-scale coding and agents that resend the same working context at every step.
What we would do
We would not migrate everything away from DeepSeek overnight, but we would stop treating it as the default low-cost lane. First, we would replay real production traffic against at least two alternatives, measuring the price of an accepted result while separately tracking cache tokens, output, failed calls and retries. Batch work can remain on V4 Flash outside peak hours; interactive and critical paths should have hard budget limits and a tested fallback provider.
Virtual Arc’s view: this is not merely a price increase; it is a warning against building agent economics around a vendor’s temporary token subsidy. DeepSeek was unusually attractive for workloads that repeatedly resend large contexts, yet cached input is where the new schedule bites hardest. We would pause any expansion of V4 Pro in production until we replay our own evaluations under the new rates. We would keep V4 Flash behind a model-routing layer, move asynchronous work outside peak windows, and measure cost per accepted task—including retries, reasoning length and human review—not cost per token. If a product works only while one provider sells inference below a durable market price, that product does not yet have durable unit economics.