On June 29, DeepSeek sent an email notification to API users: DeepSeek V4 will officially launch in mid-July 2026, accompanied by a brand-new peak-valley pricing mechanism. During peak hours, API prices double. Off-peak, baseline pricing applies.
The community reaction was immediately polarized. Some called it long overdue; others cried price gouging. But strip away the emotions, and the signal behind this pricing strategy is far more significant than the numbers themselves.
Infographic: DeepSeek-V4 peak-valley pricing overview, with CNY-denominated price comparison for V4-Pro and V4-Flash models. Source: Compiled from Yicai, 21st Century Business Herald, Guancha.cn.
The Facts: How Peak-Valley Pricing Works
According to DeepSeek's official announcement and public pricing page, the rules are straightforward:
Peak hours (all times Beijing Time, UTC+8):
- 09:00 – 12:00 and 14:00 – 18:00, weekdays
- API prices are 2× the baseline during these windows
Off-peak hours: Everything else — evenings, nights, and early mornings — at baseline price.
V4 ships in two tiers:
| Spec | DeepSeek-V4-Flash | DeepSeek-V4-Pro |
|---|---|---|
| Context length | 1M tokens | 1M tokens |
| Max output | 384K tokens | 384K tokens |
| Input (cache miss) | $0.14 / 1M | $0.435 / 1M |
| Input (cache hit) | $0.0028 / 1M | $0.003625 / 1M |
| Output | $0.28 / 1M | $0.87 / 1M |
| Concurrency limit | 2,500 | 500 |
| Thinking mode | Yes (default on) | Yes (default on) |
Notably, V4-Pro already underwent a permanent price cut during its preview phase — slashed to one-quarter of its original pricing. That means the current baseline is already deeply competitive. The peak-hour doubling is layered on top of this reduced base.
Additionally, the legacy model names deepseek-chat and deepseek-reasoner will be deprecated on 2026/07/24. They now map to V4-Flash's non-thinking and thinking modes, respectively.
Like Electricity: Not New, But a First for AI APIs
Peak-valley pricing isn't DeepSeek's invention. The electricity grid has used it for decades. Telecom operators have offered off-peak data packages for just as long. The core logic is simple: supply-side capacity is fixed, demand-side fluctuation is enormous, and price elasticity smooths the curve to maximize utilization.
But applying this logic to AI APIs? DeepSeek is the first.
For the past two years, every LLM provider used one pricing model: linear per-token rates. You pay the same whether you're batch-processing embeddings at 3 AM or serving a live chatbot during the morning rush. The implicit assumption: compute supply is abundant, and the bottleneck is customer acquisition — not resource scheduling.
V4's peak-valley pricing breaks that assumption. It signals something unmistakable: DeepSeek's GPU clusters are facing genuine peak-load pressure.
This isn't a bad thing. In fact, it's the opposite — it proves DeepSeek's API traffic has grown large enough to require capacity management. China's mobile internet went through the exact same transition when it moved from "unlimited data plans" to "time-restricted data packages." The parallel is striking.
Doubling During Peak: Is It Actually Expensive?
Let's run the numbers. For V4-Pro:
| Scenario | Input (/1M tokens) | Output (/1M tokens) |
|---|---|---|
| Off-peak baseline | $0.435 | $0.87 |
| Peak price (2×) | $0.87 | $1.74 |
Now, compare against competitors' current pricing:
| Model | Input | Output |
|---|---|---|
| GPT-4o | $2.50 | $10.00 |
| Claude Sonnet 5 (promo) | $3.00 | $10.00 |
| Claude Sonnet 5 (regular) | $3.00 | $15.00 |
| DeepSeek V4-Pro (peak) | $0.87 | $1.74 |
| DeepSeek V4-Pro (off-peak) | $0.435 | $0.87 |
Even during peak hours, V4-Pro is still only one-third to one-fifth of GPT-4o's price. Off-peak, the gap widens to 6–11×.
So no — peak doubling doesn't mean "expensive now." It means "that incredibly cheap price you've been enjoying? Going forward, you need to shift your workload to get it."
Who Wins, Who Loses
Beneficiaries:
- Batch processing and offline inference — Data labeling, bulk embeddings, model evaluation, and synthetic training data generation don't require real-time responses. Moving these tasks to nighttime cuts costs in half. The most obvious winners.
- International developers — DeepSeek's peak hours are defined in Beijing Time. For developers in Europe or the Americas, their working hours naturally fall in Beijing's off-peak window. They get the discount for free.
- Cost-sensitive startups — Flexible task scheduling to off-peak hours further compresses AI spending.
Those under pressure:
- Live customer service / real-time agents — Users won't wait until 8 PM to ask a question. Peak-hour doubling means direct cost increases with no avoidance path.
- Financial / trading real-time analysis — Market analysis and risk decisions must execute in real time. No room for off-peak scheduling.
- API proxy / relay providers — Pricing complexity increases. Customers may struggle to estimate costs when peak-hour quotes double, leading to confusion or pushback.
The Route Debate: DeepSeek Peak-Valley vs. OpenAI Batch API
It's worth noting that OpenAI and Anthropic have their own peak-shaving mechanism — the Batch API. You submit asynchronous requests, accept returns within 24 hours, and get 50% off.
Both approaches have trade-offs:
| Dimension | DeepSeek Peak-Valley | OpenAI Batch API |
|---|---|---|
| Latency | Real-time (off-peak) | Up to 24h delay |
| Predictability | Deterministic by time window | Price locked at submission |
| Flexibility | Use anytime, priced by clock | Must pre-submit, async wait |
| Discount | 50% off-peak | 50% fixed |
| Best for | Shift-able real-time tasks | Non-urgent batch processing |
DeepSeek's approach is more flexible — no need to pre-plan; just call during off-peak. OpenAI's is more controlled — price is locked at submission, immune to time-of-day fluctuations.
From a developer experience perspective, DeepSeek is friendlier (no code changes, just schedule shifts). From a platform operations perspective, OpenAI's Batch API gives tighter resource control.
The Deeper Signal: AI Compute Market Maturation
Zoom out, and DeepSeek V4's peak-valley pricing marks an industry-level inflection: the AI compute market is transitioning from "land grab" to "precision operations."
Throughout 2024–2025, the industry's theme was a price war — whoever is cheapest wins developers. DeepSeek V3 was a leader of that war, breaking through the market with extreme cost-performance.
But price wars can't last forever. When API call volume grows past a threshold, the supply side must start managing resource efficiency. Peak-valley pricing is the first step of post-price-war refinement.
This trajectory closely mirrors China's cloud computing market. Alibaba Cloud and Tencent Cloud started with "¥9.9/year" customer acquisition, then evolved to on-demand billing, reserved instances, and spot instances. AI APIs are walking the same path.
We can expect more providers to follow suit:
- Time-based pricing will become standard (peak, off-peak, and possibly shoulder tiers)
- Priority-based pricing may emerge (high-priority requests cost more, low-priority costs less)
- Spot instances could appear (idle capacity auctioned, similar to AWS Spot Instances)
DeepSeek is the first to eat the crab. Whether it tastes good depends on V4's actual performance after launch.
Practical Advice for Developers
If you're a heavy DeepSeek API user, here are strategies for the peak-valley era:
- Audit your call logs — Measure what fraction of token consumption falls in peak hours. If more than 50% of calls can be shifted, the savings are significant.
- Add cron jobs for batch tasks — Schedule non-real-time workloads (data annotation, embeddings generation, eval suites) after 18:00 or before 09:00 Beijing Time.
- Cache hit rate is the new optimization frontier — V4's cache-hit price is 1/50th of cache-miss. This means prompt engineering for prefix consistency is now 50× more valuable. Small investments in prompt structure yield outsized returns.
- Relay services need repricing — If you operate an API proxy, your pricing model must reflect peak-off-peak differences. Otherwise, peak-hour margins get eaten alive.
Final Thoughts
The controversy around DeepSeek V4's peak-valley pricing reflects a fundamental shift in how the AI industry thinks about compute: it's not air — it shouldn't be free; but it's not a luxury good either — it needs to be allocated efficiently.
The electricity industry took 100 years to move from "flat rate" to "time-of-use pricing." AI compute did it in two.
This isn't just one company's pricing adjustment. It's a rite of passage as an industry moves from adolescence to maturity. Whether V4's performance lives up to expectations — we'll see in mid-July. But peak-valley pricing? That's already written into the history of AI infrastructure.