DeepSeek ends the cheap-token era with peak and off-peak API pricing
The lab that set the floor for agent inference costs now charges by the clock. Peak-hour output for V4 Pro runs roughly four times the old promotional rate, and off-peak hours cost half of peak.
DeepSeek will start charging by the clock. The Chinese AI lab said its V4 API moves to peak and off-peak pricing at 16:00 UTC on Sunday, the start of Monday in Beijing, replacing the flat promotional rates that made it the reference point for cheap agent inference. TechNode reported that peak hours run from 9 a.m. to noon and from 2 p.m. to 6 p.m. Beijing time, with all other hours billed at half the peak rate.
The size of the increase is the headline. According to Engadget, output tokens for the flagship V4 Pro rise to $3.96 per million during peak hours, up from $0.87 under the old rates, while the smaller V4 Flash climbs to $1.32 from $0.28. Off-peak, those figures drop to $1.98 and $0.66. DeepSeek said the new structure is meant "to allocate resources more reasonably," language that suggests its clusters are saturated during the Chinese business day and idle overnight.
The change also closes out a long flip-flop. The discounted rates began as a promotion scheduled to end May 31. DeepSeek then announced the promotional prices would become permanent, a move that pressured rivals through the spring. This weekend's update scraps that commitment entirely. The contrast with Anthropic, which converted its own promotional rates into permanent Sonnet 5 pricing last week, will not be lost on buyers deciding where to anchor long-lived agent deployments.
For agentic AI services, the mechanics matter more than the increase. Time-of-day pricing is a utility model, and it maps cleanly onto how agent workloads actually run. Batch jobs such as overnight document processing, code migrations, and evaluation sweeps can be scheduled into off-peak windows at half price. Latency-sensitive customer-facing agents pay full freight. Orchestration layers that already route between models on price will now have a second variable to optimize, and operators serving customers in the United States and Europe happen to sit largely outside DeepSeek's Beijing-daytime peak window.
Even after the increase, DeepSeek undercuts most Western rivals. Engadget notes that OpenAI's GPT-5.6 Sol runs $30 per million output tokens and Moonshot's Kimi K3 about $15, figures that leave V4 Pro at a fraction of frontier pricing even at peak. But raw token rates were never the whole story for agent economics. As GaaS News has reported, long tool loops multiply whatever the meter charges, and a fourfold jump at the base of the stack compounds through every retry and reflection step. Agent operators who built unit economics on DeepSeek's promotional floor have until Sunday evening to re-run the math.