Skip to main content

DeepSeek V4 Goes Live in Mid-July: Peak-Hours API Pricing Doubles, on Top of a Permanent Cut

On June 29, 2026, DeepSeek sent an upgrade-reminder email to a subset of users, bundling two announcements: the mid-July launch of V4 proper and a time-of-use (peak/off-peak) pricing mechanism. The email's details were first disclosed by Guancha (Observer) on June 30. What developers should actually pay attention to is not the surface-level "price increase", but the strategic pivot behind it — DeepSeek's pricing philosophy is moving from "uniformly cheap" to "precision cheap".

1. The Two Moves in the Email​

According to the email content disclosed by Guancha, DeepSeek released two pieces of information simultaneously:

  1. The V4 official release is scheduled for mid-July, accompanied by more feature optimizations and performance improvements.
  2. A time-of-use pricing mechanism is being introduced simultaneously: Beijing time 9:00–12:00 and 14:00–18:00 are designated as peak hours, with call prices set at 2x the off-peak rate; all other hours are billed at the off-peak rate.

The email also promises that any related adjustment will be notified to users via email 24 hours in advance. This is a very engineering-friendly buffer clause — it gives operations a full business day to complete traffic shifting, cache warm-up, or rate-limit reconfiguration.

2. The Real Price Table With Peak Hours​

Layering the peak-hour doubling on top of V4's existing pricing yields 4 time-slots × 3 cost tiers = 12 numbers. Sorted out below (per million tokens, USD; the original email used RMB, converted at ~1 RMB ≈ $0.14):

SlotModelInput (cache hit)Input (cache miss)Output
Off-peakV4-Flash$0.0028$0.14$0.28
Off-peakV4-Pro$0.003625$0.435$0.87
PeakV4-Flash$0.0056$0.28$0.56
PeakV4-Pro$0.00725$0.87$1.74

Compared to V4-Pro's pre-cut ceiling of 24 RMB per million tokens (≈ $3.36) before the May 23 permanent price drop, V4-Pro's output price — even with the peak-hour doubling — is still only half of that historical high. In other words, the "price increase" is relative to the May-end sweet spot, not relative to the V4 Preview pricing published in April.

3. Real Impact on AI Apps: V4-Pro Output as the Worked Example​

Guancha published a very concrete cost comparison: an AI application consuming 100 million output tokens per day during peak hours pays roughly 1,200 RMB at peak rates, versus ~600 RMB at off-peak rates; if consumption climbs to 1 billion output tokens/day, the cost moves from ~6,000 RMB to ~12,000 RMB. For customer service, code assistants, office agents, and search-augmented Q&A — the canonical high-traffic-during-peak pattern — the doubling directly compresses gross margin.

This implies three things:

  • Peak-avoidance design is now mandatory: deferrable batch jobs, offline analysis, and async-retry-able conversations must be pushed past 18:00.
  • Prompt length must be re-evaluated: at peak hours, every additional 1K input tokens costs an extra $0.00725 per million (V4-Pro cache hit) — a level of granularity that simply did not register in the V3 era.
  • Context Caching must become a default capability: cache hit and miss tiers are both doubled at peak, but the absolute gap between the two tiers is also widened — meaning the marginal ROI of caching in V4 is higher than in V3.

4. International Comparison: Still a "Smoke" Deal​

Zooming out to global pricing, DeepSeek's confidence in raising prices is well-grounded. At peak rates, V4-Pro output lands at roughly $1.74 per million tokens, versus OpenAI's published GPT-5.5 standard API pricing (input $5, cache input $0.50, output $30 per million) and Anthropic's Claude Opus 4.8 standard pricing (input $5, output $25 per million). OpenAI and Anthropic high-end output prices are still about 14–17x DeepSeek V4-Pro's peak-hour rate.

That is the real moat behind DeepSeek's July "price increase": it is not testing user tolerance for higher prices, but reserving ample competitive headroom via the international price gap. For budget-constrained overseas users, DeepSeek's peak pricing still sits at the low end of the global price anchor.

5. Context: From "Permanent Cut Before Increase" to "Precision Cheap"​

This is not DeepSeek's first pricing move in 2026. The timeline:

  • April 24: V4 Preview launched, with V4-Pro (1.6T/49B) and V4-Flash (284B/13B) — both with 1M context, dual concurrency caps (V4-Flash 2,500 / V4-Pro 500).
  • May 23: V4-Pro's 75% promotional discount was converted into a permanent price, with the API ceiling dropping from 24 RMB to 6 RMB per million tokens.
  • June 29: Email previews mid-July V4 proper + time-of-use pricing.

The cadence is exceptionally clear: lock in power users with Preview, broaden the ecosystem with a permanent cut, then explicitly price the scarcity of compute via peak/off-peak windows. This is not abandoning the low-price strategy; it is upgrading "uniformly cheap" to "time-stratified cheap".

6. Industry Trend: Overseas Enterprises Are Being Re-educated by "Multi-Model Routing"​

Guancha also highlights the macro context behind DeepSeek's price increase:

  • In recent months, overseas LLM pricing has shifted from flat subscriptions to per-token billing, causing enterprise usage costs to climb rapidly.
  • Uber burned through its full-year AI budget in 4 months, forcing executive-usage restrictions — it has been called "the first big tech company to halt AI cash-burn".
  • Executives at Microsoft, Coinbase and others have begun emphasizing that "many tasks don't always need the most expensive, largest models".
  • OpenRouter data shows that open-source models now handle ~65% of token throughput on the platform, with usage of Chinese low-cost models like DeepSeek visibly rising.

In other words, DeepSeek is rolling out time-of-use pricing right inside the global "AI budget tightening" window. Overseas customers are actively reworking their routing strategies — handing simple tasks to cheap models and reserving premium models for the hard ones — and DeepSeek is packaging "cheap" as a time-quantifiable product.

7. An Action Checklist for Developers​

Translating the above into engineering actions before mid-July:

  1. Build a "peak traffic map": use 2–4 weeks of real logs to draw request density and token consumption per endpoint during the 9–12 and 14–18 windows, and identify deferrable batch jobs.
  2. Reschedule task execution: force-run ETL, long-document summarization, batch classification, and vector recomputation into the post-18:00 off-peak window.
  3. Make Context Caching a default capability: peak hours widen the absolute gap between cache hit and miss, so the ROI of caching rises sharply — push the hit-rate target from 30% to 60%.
  4. Run a billing replay before 7/24: replay the last 30 days of real traffic through the new price table, quantify the "peak increase" per SKU, and set monthly budget alerts per line item.
  5. Watch the official DeepSeek email: the email promises 24-hour advance notice — add api-service@deepseek.com to your IM alerting pipeline so you never react late.

8. Closing Thoughts​

By bundling V4 proper and time-of-use pricing for a mid-July launch, DeepSeek is, at its core, upgrading its pricing philosophy: from "uniformly cheap" to "time-stratified cheap". The peak-hour doubling sounds painful, but the 14–17x international gap still holds. For developers, the real homework for mid-July is not "complain about the price increase" but redesign traffic scheduling: which tasks must run at peak, which can be shifted to off-peak, and which can dodge the doubling via cache hits. The answers will directly determine next quarter's API bill.

Sources: Guancha (Observer) exclusive report dated 2026/06/30 (m.guancha.cn/economy/2026_06_30_822053.shtml), the DeepSeek June 29 upgrade-reminder email, and DeepSeek's official API documentation (api-docs.deepseek.com/quick_start/pricing). Written based on public information as of 2026/07/02. The final details of V4 proper and time-of-use pricing are subject to DeepSeek's official July announcement.