TL;DR — One-Sentence Conclusion
On June 29, 2026, the DeepSeek team announced that DeepSeek-V4 official release will launch in mid-July, debuting a first-ever “peak-valley pricing” mechanism for LLM APIs—rates double during peak hours (daily 09:00-12:00 + 14:00-18:00, ~7 hours), while the remaining 17 hours per day keep current pricing. This is the first time the electricity-style “Time-of-Use Tariff” logic has been ported into LLM API billing, shifting the industry mindset from “which model is cheapest” to “what time of day to call it”.
Sources: DeepSeek official announcement 2026-06-29, Tencent Cloud notice 2026-07-03, Caixin/Beike Finance 2026-06-29, html5.qq.com 2026-06-30
1. DeepSeek-V4 Official Release Timeline: One Week of Major Milestones (6/29-7/3)
| Date | Event | Source |
|---|---|---|
| 2026-04-24 | V4 Preview released + open-sourced (1.6T total params / 49B active / 1M context) | DeepSeek WeChat official |
| 2026-05-22 | V4-Pro permanent price cut (1/5.33 of original; ~8× cheaper than Claude Opus 4.7) | CSDN 2026-05-30 |
| 2026-06-29 | DeepSeek announces V4 official release in mid-July + peak-valley pricing | Beike Finance / Sina Tech |
| 2026-06-30 | Pricing details disclosed: V4-Pro peak 12 RMB/M, Flash peak 4 RMB/M | html5.qq.com (4-source cross-check) |
| 2026-07-03 | Tencent Cloud: TokenHub + Agent platform to follow the pricing strategy | Tencent Cloud official notice |
2. DeepSeek-V4 Core Specs at a Glance
| Spec | V4-Pro | V4-Flash | Benchmark Comparison |
|---|---|---|---|
| Total parameters | 1.6 trillion | 284 billion | vs GPT-5.4 / Claude Opus 4.6 |
| Active parameters | 49 billion | 13 billion | same as above |
| Pre-training data | 33T tokens | 32T tokens | same as above |
| Context window | 1M tokens | 1M tokens | 1M+ becoming industry standard |
| Coding capability | SWE-Bench global top tier | Beats Pro in 5/20 real tasks | Surpasses Claude Opus 4.6 |
Key technical breakthrough: Engram Conditional Memory—separates “fact storage” from “reasoning” via an external retrievable knowledge base with O(1) constant-time lookup. At 1M-token context, V4-Pro uses only 27% of V3’s per-token inference compute and 10% of V3’s KV-cache footprint (CSDN 2026-04-27 deep dive).
Sources: DeepSeek official technical report 2026-04-24, CSDN 2026-04-27 “Million-Token Context for the Masses”, CSDN 2026-07-01 “V4 Comprehensively Beats Claude Opus 4.6”
3. Peak-Valley Pricing: 7 Hours Doubled, 17 Hours Unchanged
Peak hours: daily 09:00-12:00 + 14:00-18:00 (combined 7 hours/day), covering the prime “office-hour work block” of most enterprises.
V4-Pro price table (per million tokens / RMB):
| Billing item | Off-peak | Peak | Multiplier |
|---|---|---|---|
| Input (cache hit) | 0.025 | 0.05 | 2× |
| Input (cache miss) | 3 | 6 | 2× |
| Output | 6 | 12 | 2× |
V4-Flash price table:
| Billing item | Off-peak | Peak | Multiplier |
|---|---|---|---|
| Input (cache hit) | 0.02 | 0.04 | 2× |
| Input (cache miss) | 1 | 2 | 2× |
| Output | 2 | 4 | 2× |
Head-to-head with international flagships (5/29 USD/M token prices):
- GPT-5.4 (OpenAI): Input 2.50 / Output 15.00 (1.1M context)
- Claude Opus 4.6 (Anthropic): 5.00 / 25.00 (1.0M)
- Gemini 3.1 Pro Preview (Google): 2.00 / 12.00 (1.0M)
- DeepSeek-V4-Pro off-peak: ~USD 0.35 / 0.84 (converted at 7.16 RMB/USD) — approximately 14-30× cheaper than Claude Opus 4.6
Sources: DeepSeek official 2026-06-29, Tencent Cloud 2026-07-03, CSDN 2026-05-29 “2026 Complete LLM API Pricing/Speed/Chinese-Capability Comparison”
4. Five Deep Implications for Enterprise AI Deployment
Signal 1: Time-of-Use Tariff Era Opens — “When to Call” Beats “Which Model” Previously, enterprises just compared per-token prices across DeepSeek/Claude/Qwen. Now they also need to ask: can my batch inference run after 8 PM? This means AI engineers are starting to become power-grid dispatchers.
Signal 2: Open-Source Flagship Is Now an Order of Magnitude Cheaper Than Closed-Source Converted at off-peak rates, V4-Pro is roughly 14-30× cheaper than Claude Opus 4.6. This price gap pushes a dual-track pattern: DeepSeek for China SaaS / overseas markets, closed-source for premium overseas.
Signal 3: “Cheap but Capable” Flash Logic Is Battle-Tested Independent testing showed Flash beats Pro in 5 of 20 real-world tasks—lightweight versions genuinely replace flagships in many scenarios. This elevates “fast/slow twin-track” from architectural choice to enterprise cost-saving default.
Signal 4: Tencent Cloud, Alibaba Cloud and Other Majors Actively Follow Pricing Strategy Tencent Cloud’s 7/3 notice used the explicit phrase “follow the manufacturer synchronously”—LLM APIs evolved from “vendor unilateral pricing” to “ecosystem collective bargaining”. Any single vendor’s price move will now be market-tracked.
Signal 5: “Coding + Agent” Becomes the Battleground Where Open Source Overtakes Closed Source V4-Pro’s Agentic Coding real-world experience is better than Claude Sonnet 4.5, with delivery quality near Opus 4.6 non-thinking mode (CSDN 2026-04-27)—“Chinese AI coding surpasses US” moves from slogan to benchmark fact. For AI deployment service providers like Xi’an Boao, this is a direct upwind signal.
5. One Action Item for Xi’an Boao AI
Boao AI’s current client-facing assistants mostly rely on OpenAI/Anthropic, with API costs running 15-25% of project margins. V4-Pro at off-peak pricing is 1/30 the cost of Claude Opus 4.6. Recommend that after the mid-July official release:
- Phase 1 (within 2 weeks): Migrate all off-peak batch tasks (data cleansing, report generation, knowledge-base vectorization) to V4-Flash off-peak.
- Phase 2 (1-2 months): Migrate the dialogue main path to V4-Pro off-peak, keeping Sonnet as the “peak-hour emergency fallback.”
- Client communication: Add one line in July client reports: “Now operating in time-of-use tariff mode”—a showcase of Boao’s technical acuity.
FAQ (High-Frequency Questions)
Q1: Why is peak defined as 09:00-12:00 + 14:00-18:00? A: These 7 hours cover peak enterprise usage, using price signals to shave demand off peak hours and fill valleys at night/early morning, improving resource utilization.
Q2: Is off-peak pricing cheaper than current? A: No, same as current. V4-Pro/Flash off-peak rates match the 5/22 permanent price cut—only peak hours rise above current.
Q3: Will peak-valley pricing make enterprises abandon international vendors entirely? A: No. Night-time batch jobs enjoy cheap off-peak rates, but daytime real-time interaction still demands upper-bound model capability—closed-source flagships retain a 5-15% edge in code reasoning and long-doc understanding.
Q4: Are free open-source users affected? A: Not directly. The official chat interface at chat.deepseek.com stays free—peak-valley pricing only applies to paid API calls.
Q5: After Tencent Cloud TokenHub follows, will other clouds (Alibaba, Huawei, Volcano) follow? A: Highly likely. The phrase “follow the manufacturer” in Tencent Cloud’s notice already signals ecosystem-aligned pricing—a landmark event marking LLM APIs’ shift from “vendor unilateral pricing” to “platform collective negotiation”.
Q6: What will V4 official improve vs the 4/24 preview? A: DeepSeek only stated “feature optimization and performance improvement” without concrete metrics. Third-party predictions focus on: (1) multimodal gap-filling (V4-Pro is text-only, missing native vision/speech); (2) Agent tool-call stability; (3) ultra-long context extension (1M+ → 4M).
Key Terminology
- Peak-Valley Pricing (Time-of-Use Tariff): Originated in the electricity industry—a tariff mechanism that charges different rates by usage time window to flatten demand. DeepSeek is the first to port it into LLM API billing.
- V4-Pro / V4-Flash: DeepSeek’s two-version flagship launched 4/24. Pro targets closed-source top tier; Flash focuses on “low cost + high response.”
- 1M tokens context: A single conversation can hold ~750K Chinese characters or ~500K English words—roughly the size of a medium-thickness book.
- Engram Conditional Memory: V4’s core architectural innovation that decouples knowledge storage from reasoning—the key to keeping 1M-token context low-latency.
- SWE-Bench: Real-world software engineering benchmark measuring model ability to fix bugs on actual GitHub issues—the empirical yardstick for coding capability.
- Cache Hit: When a request reuses prior KV cache (because of similar content), the price is roughly an order of magnitude lower than cache miss.
- Agentic Coding: Letting AI autonomously call tools, write code, run tests, and fix bugs—a dimension above simple code completion.
References
Official Announcements
- DeepSeek Official: “DeepSeek-V4 Official Release Mid-July” (2026-06-29)
- Tencent Cloud: “Notice on DeepSeek-V4 Official ‘Direct-Supply’ Model Release Plan and Pricing Adjustment” (2026-07-03)
Industry Coverage
- Beike Finance: “DeepSeek V4 Official Release Mid-July, Introducing Peak-Valley Pricing” (2026-06-29)
- Xiaoxiang Morning Post / Tencent News: “Tencent Cloud Announces DeepSeek-V4 ‘Direct-Supply’ Model Mid-July Release Plan” (2026-07-03)
- html5.qq.com: “DeepSeek V4 Official Confirmed Mid-July” (2026-07-02)
- html5.qq.com: “Peak-Hour Prices Double! DeepSeek V4 Official Confirmed” (2026-07-01)
Technical Analysis
- CSDN: “DeepSeek-V4 Preview Released: Million-Token Context for the Masses” (2026-04-27)
- CSDN: “DeepSeek V4 Storms In: Open-Source Model Performance Reversal, Comprehensively Beats Claude Opus 4.6” (2026-06-30)
- CSDN: “DeepSeek V4-Pro Permanent Price Cut and Composer 2.5 Launch” (2026-07-02)
Pricing Comparison & Background
- CSDN: “2026 Complete LLM API Price / Speed / Chinese-Capability Comparison (May Update)” (2026-05-29)
- SegmentFault: “Complete LLM Evaluation Guide: 2026 Mainstream LLM Benchmarks” (2026-03-11)
- 36Kr: “Will ‘LLM Price War’ Keep Burning Money in 2025?” (2025-01-16)