⚠️ Important Update: Effective August 17, 2026, DeepSeek implemented peak/off-peak pricing. V4-Pro and V4-Flash rates changed significantly. Peak hours (weekdays 9:00-12:00, 14:00-18:00 Beijing time) cost 2x off-peak rates. This article reflects the latest pricing.
Alibaba Cloud Bailian · DeepSeek Access
Enterprise-grade SLA with free credits for new users
📰 Latest Update (Oct 6, 2026)
DeepSeek Harness v0.2.1 ships official desktop apps + Claude Code Mods compatibility: On Oct 3, DeepSeek released dsh v0.2.1 (MIT, ~243K GitHub stars) with Windows x64 / macOS desktop apps and an experimental Claude Code Mods compatibility layer — validating that Mods APIs are broadly a subset of dsh’s everything-is-a-plugin architecture. The self-hosted, plugin-first coding-agent ecosystem keeps expanding, making open-source agents an increasingly cost-effective alternative for teams running DeepSeek APIs in production.
DeepSeek publishes DSec, its sandbox infrastructure for large-scale agentic training: On Sep 23, DeepSeek released the system paper "DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale" (arXiv 2609.22978, 31 pages, 100+ authors including Liang Wenfeng), detailing its production agent-training platform: one production cell packs ~160 CPU nodes, 30,000 CPU cores and 250TB DRAM, serving ~3 million sandboxes per day with peak concurrency of 380,000+ and creation speed of 5,000+ sandboxes/second; all RL training and evaluation from DeepSeek V3.2 to V4.1 has run on DSec. On-demand image loading cuts 8,192-container startup time by ~42%. As agentic training scales demand for CPU, memory and sandbox capacity, elastic on-demand compute remains the key to controlling costs.
1. Latest DeepSeek Pricing (Effective Aug 17, 2026)
DeepSeek now offers three main model versions with time-based pricing:
| Model | Billing Item | Off-Peak | Peak | Unit |
|---|---|---|---|---|
| V4-Pro (Flagship) | Input (cache miss) | $0.63 | $1.25 | per 1M tokens |
| Input (cache hit) | $0.02 | $0.04 | per 1M tokens | |
| Output | $1.88 | $3.75 | per 1M tokens | |
| V4-Flash (Fast/Light) | Input (cache miss) | $0.21 | $0.42 | per 1M tokens |
| Input (cache hit) | $0.007 | $0.014 | per 1M tokens | |
| Output | $0.63 | $1.25 | per 1M tokens | |
| V3 (Stable) | Input/Output | Input $0.14 / Output $0.56 (flat rate) | per 1M tokens | |
Peak hours: Weekdays 9:00-12:00 and 14:00-18:00 (Beijing time, UTC+8). All other 17 hours on weekdays, plus weekends and public holidays, are off-peak (50% of peak price). Prices shown in USD at approximate 7.2 CNY/USD rate.
2. Platform Comparison
Below is a comparison of major platforms offering DeepSeek-V3 API access:
| Platform | Model | Input | Output | SLA | Rating |
|---|---|---|---|---|---|
| Alibaba Cloud Bailian | DeepSeek-V3 | $0.14/1M | $0.56/1M | 99.9% | ★★★★★ |
| DeepSeek-R1 | $0.56/1M | $2.22/1M | |||
| Volcano Engine | DeepSeek-V3 | $0.14/1M | $0.56/1M | 99.9% | ★★★★☆ |
| SiliconFlow | DeepSeek-V3 | $0.13/1M | $0.53/1M | 99.5% | ★★★★☆ |
| DeepSeek Official | DeepSeek-V3 | $0.14/1M | $0.56/1M | None | ★★★☆☆ |
Alibaba Cloud Bailian DeepSeek-R1 pricing: Input $0.56/1M tokens, output $2.22/1M tokens (Beijing region). R1 excels at mathematical reasoning, coding, and logical inference tasks.
3. Cost-Saving Strategies with Peak/Off-Peak Pricing
- Schedule batch jobs off-peak: Data analysis, document generation, and batch translations running between 18:00-9:00 Beijing time or on weekends cost 50% less.
- Leverage prompt caching: V4-Pro cache-hit input costs only $0.02/1M off-peak vs $0.63 for cache misses — a 30x difference. Applications with stable system prompts benefit enormously.
- Choose the right model: Use V4-Flash for everyday tasks (off-peak output $0.63/1M), reserve V4-Pro for complex reasoning. V3 remains flat-rate for predictable budgeting.
- Pre-paid resource packages: Alibaba Cloud Bailian offers resource bundles that can reduce costs by ~30% for high-volume users.
4. Enterprise Selection Criteria
- SLA guarantee: Production environments require 99.9%+ uptime commitments. Alibaba Cloud Bailian and Volcano Engine both meet this standard.
- Compliance: Finance, healthcare, and government sectors need Level 3 Cybersecurity Classified Protection. Alibaba Cloud holds Level 3, SOC2, and ISO27001 certifications.
- Ecosystem integration: Teams already using Alibaba Cloud ECS, OSS, or RDS benefit from intranet API calls with latency as low as 30ms and reduced data transfer fees.
- Technical support: Enterprise users should select platforms offering 24/7 support with rapid response times.
Start Using DeepSeek API
New Alibaba Cloud Bailian users receive free token credits
FAQ
How are peak/off-peak hours defined?
Peak: weekdays 9:00-12:00 and 14:00-18:00 Beijing time (UTC+8). Off-peak: all other hours on weekdays, plus all weekends and Chinese public holidays. Off-peak rates are 50% of peak rates.
What is the difference between V3 and V4?
V3 is the stable version with flat-rate pricing ($0.14 input / $0.56 output per 1M tokens). V4-Pro is the flagship reasoning model with peak/off-peak pricing. V4-Flash is the lightweight, cost-effective option. Choose based on your task complexity and budget.
What is prompt caching?
When multiple requests share the same prefix (such as a system prompt), DeepSeek caches that content. Subsequent requests pay the cache-hit rate, which is dramatically cheaper. V4-Pro cache-hit input is only $0.02/1M off-peak.
How can I further reduce API costs?
1) Run non-urgent batch jobs during off-peak hours; 2) Purchase pre-paid resource packages; 3) Optimize prompts to maximize cache hits; 4) Use Flash instead of Pro for simple tasks; 5) Purchase through authorized channels for additional discounts.