DeepSeek Retires Legacy Aliases and Introduces Utility-Style Peak-Hour API Pricing
Chinese AI lab DeepSeek is introducing a fundamental shift in how developer API access is priced. Alongside a mandatory migration deadline of July 24, 2026 (at 15:59 UTC) to retire its legacy deepseek-chat and deepseek-reasoner aliases, the company is launching the "Full-Power Version" of its DeepSeek V4 model family. With this general availability (GA) release, DeepSeek is introducing a novel "Peak-Valley" (Peak-Off-Peak) pricing mechanism, a first for the global AI developer market.
The DeepSeek V4 GA Launch
The official V4 release consists of two main tiers:
- DeepSeek V4 Pro: The flagship Mixture-of-Experts (MoE) reasoning model. It features trillions of parameters, a 1-million-token context window, and SWE-bench Verified coding performance that matches Anthropic's Claude Opus 4.6 Max and rivals OpenAI's GPT-5.6 Sol.
- DeepSeek V4 Flash: A lighter, highly optimized model designed for cost-sensitive, high-throughput production workloads.
The Peak-Valley Pricing Structure
Rather than maintaining a flat rate, DeepSeek's new utility-style pricing model charges developers double the standard rate during peak traffic hours, while offering massive discounts during off-peak windows1:
-
DeepSeek V4 Pro Pricing:
- Off-Peak Output: $0.87 per million tokens.
- Peak Output: $1.74 per million tokens.
- Off-Peak Cache-Missed Input: $0.435 per million tokens.
-
DeepSeek V4 Flash Pricing:
- Off-Peak Output: $0.28 per million tokens.
- Peak Output: $0.56 per million tokens.
- Cache-Hit Input: $0.0028 per million tokens (extremely subsidized).
For batch tasks, benchmark testing, and data generation workloads that can be scheduled during non-peak hours, DeepSeek remains the undisputed "price slasher" of the industry. When compared to the flat $50 per million output tokens charged for Anthropic's Claude Fable 5, DeepSeek V4 Flash represents a roughly 137x price spread for competitive agentic workloads.
Mandatory API Migration
Developers must migrate their integrations before July 24, 2026, at 15:59 UTC. After this deadline, the legacy aliases deepseek-chat and deepseek-reasoner will be permanently retired. The migration is straightforward: developers must transition to the deepseek-v4-flash model name and explicitly configure the thinking mode via DeepSeek's API parameters.
-
An instance of Corporations must now treat AI compute as a volatile, utility-priced commodity rather than a fixed infrastructure cost. — DeepSeek's Peak-Valley pricing structure establishes utility commodity mechanics where developer tasks must be optimized around volatile peak computing periods. ↩︎