DeepSeek V4: Legacy Aliases Gone, Peak Pricing Announced but Not Active
DeepSeek's V4 models now require explicit Flash/Pro choice; surge pricing is coming but not yet enforced.
DeepSeek officially retired the legacy API aliases deepseek-chat and deepseek-reasoner on July 24, 2026, as planned. Any code still routing requests to those model names now receives HTTP errors. The migration requires developers to update endpoints to the new V4 model IDs: deepseek-v4-flash for high-speed, cost-efficient workloads (chat, extraction, classification) and deepseek-v4-pro for complex reasoning, instruction-following, and long-context tasks up to 1M tokens. The API remains OpenAI-compatible at https://api.deepseek.com, so base URL and request shape are unchanged.
Meanwhile, a peak-hour surcharge tied to Beijing business hours (08:30–00:30 UTC+8) was announced but has not gone live. No surcharge percentage or start date has been published, and billing continues on a single flat per-million-token rate per model. Developers do not need to reschedule batch jobs yet, but should instrument requests with timestamps to future-proof cost attribution. DeepSeek V4 Pro competes with GPT-5.5 and Claude Opus 4.8 on reasoning benchmarks, while V4 Flash remains one of the cheapest capable models. The migration checklist emphasizes endpoint updates, graceful error handling, and logging for peak/off-peak tracking.
- Legacy aliases deepseek-chat and deepseek-reasoner retired July 24, 2026; calls now return HTTP errors.
- New model IDs: deepseek-v4-flash for general tasks, deepseek-v4-pro for complex reasoning up to 1M tokens.
- Peak-hour surcharge announced for Beijing business hours but not yet active; no percentage or start date.
Why It Matters
Developers must update API endpoints now while monitoring for upcoming pricing changes that could affect costs.