Pricing Watch｜DeepSeek V4 Hikes Prices Across the Board, Peak Hours Up to 1,100%

🌏 中文版

Change Summary

Alongside the GA release of V4-Pro, DeepSeek announced a sweeping price increase across the entire V4 API lineup and introduced peak/off-peak dual-rate billing for the first time — the most significant price hike since it adopted its low-price strategy in 2025. The most dramatic line item: V4-Pro Cache Hit Input during peak hours went from $0.003625 to $0.044/1M tokens, a 1,114% increase — the figure behind the "prices up over 1,100%" headlines. But what actually hits most applications' budgets is Output pricing, which rose 355% (V4-Pro) to 371% (V4-Flash). Even after the hike, DeepSeek remains cheaper than GPT-5.6 and Claude, but the narrative of "Chinese models at near-zero cost" is officially over.

Before & After Comparison

Off-peak rates are fixed at 50% of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; all other hours are off-peak.

Cost Estimate

Scenario: An Agent handling 10,000 customer service conversations per day (averaging 1,500 input tokens + 500 output tokens each, no caching), running on DeepSeek V4-Pro.

For the same workload, simply shifting traffic to off-peak hours saves $594/month — more than the old total bill. This change makes "when you call the API" as important a cost decision as "which model you call."

Impact on Developers & Enterprises

Who Benefits Most

Nobody "benefits" from a price hike in the short term, but teams that can schedule non-real-time tasks into off-peak windows are least affected. Batch summarization, data cleaning, offline eval runs — anything that doesn't need instant responses can move to off-peak and nearly halve the increase. Conversely, real-time customer service and chatbot traffic that can't be scheduled will absorb the full 355%+ hike during peak hours.

Competitive Landscape Impact

Post-hike Output pricing rankings (USD/1M tokens) for major models:

Even after the hike, DeepSeek's peak prices land near Claude Haiku 4.5, and off-peak prices remain well below all major Western models — this adjustment narrowed the gap without flipping the rankings.

Recommended Actions

• If your workload tolerates latency (batch processing, summarization, offline RAG index builds): shift scheduling outside the 01:00–04:00 / 06:00–10:00 UTC peak windows for an automatic 50% discount.
• If you're running real-time customer service or conversational Agents: re-run your cost projections at peak prices. The 355%-371% increase may make the DeepSeek vs. Claude Haiku 4.5 spread negligible — worth re-evaluating quality and latency before deciding whether to switch.
• If your system relies heavily on prompt caching: note that Cache Hit Input saw the largest hike (up to 1,114% at peak). The assumption that cache hits are "practically free" needs revisiting — don't estimate total cost based on Output increases alone.

Takeaway

The media headline "prices up over 1,100%" refers to V4-Pro peak-hour Cache Hit Input — the smallest, least commonly factored line item in total cost. The increase most applications actually feel (Output, 355%-371%) is far less dramatic than the headline number. When reading pricing news, always verify "which line item increased" first — headlines tend to pick the cell with the smallest absolute value and the most eye-catching percentage.

References

• DeepSeek-V4-Pro GA Release
• Models & Pricing | DeepSeek API Docs
• DeepSeek raises some V4 prices by more than 10x as AI demand strains capacity - InfoWorld
• DeepSeek Increases Prices for AI Services by Multiple Times - Bloomberg (via Yahoo Finance)
• DeepSeek Launches V4-Pro and Raises API Prices by as Much as 1,100% - Caixin Global
• DeepSeek's AI Models Are About To Cost Four Times More - Engadget