DeepSeek V4 Migration: Legacy API Aliases Retired July 24 — and Where Surge Pricing Actually Stands

1. Hard Deadline: Legacy Aliases Retire July 24

If your codebases, SDK configs or environment variables still route requests to deepseek-chat or deepseek-reasoner, migrate now. DeepSeek has been silently forwarding those aliases to V4 Flash under the hood, but backward compatibility ends on July 24, 2026. After that, calls to legacy model names return HTTP errors.

Update your endpoints directly to the V4 model IDs:

• deepseek-v4-flash — high-speed, cost-efficient default for most chat, extraction and classification workloads.
• deepseek-v4-pro — complex reasoning, heavy instruction-following and long-context work up to 1M tokens.

Base URL and OpenAI-compatible request shape are unchanged (https://api.deepseek.com). See our DeepSeek API guide and pricing page for full rate cards.

2. Time-of-Day Surge Pricing: Announced, Not Active

The change that got the most attention was a peak-hour surcharge tied to Beijing business hours (UTC+8). It is worth being precise about its status, because a lot of coverage reported it as live:

• Announced — DeepSeek has said a peak-hour surcharge is coming, with the peak window framed around Beijing daytime (08:30–00:30 UTC+8).
• Not in effect — no surcharge percentage and no start date have been published, and billing today runs on one flat per-million-token rate per model.

Historically DeepSeek ran off-peak discounts; that programme has ended. The announced surcharge inverts the mechanism — same load-balancing goal, opposite framing — but until it ships there is no time-of-day component in a DeepSeek invoice.

3. V4 Performance: Holding the Price-to-Performance Crown

DeepSeek built its reputation by delivering frontier-adjacent quality at a fraction of Western API prices. V4 doubles down: V4-Pro competes with GPT-5.5 and Claude Opus 4.8 on reasoning benchmarks, and V4-Flash remains one of the cheapest capable models on the market. Even as open-weight rivals like Kimi K3 push parameter counts higher, DeepSeek V4 keeps its edge on the speed × capability × cost triangle.

For a deeper look at how DeepSeek is squeezing more out of every GPU, see our write-ups on DSpark speculative decoding and DeepSeek's custom inference chip.

Developer Migration Checklist

• Update endpoints: Audit code and swap deepseek-chat / deepseek-reasoner for deepseek-v4-flash or deepseek-v4-pro.
• Do not reschedule yet: there is no time-of-day rate today, so moving cron jobs buys nothing until the announced surcharge actually ships.
• Verify fallback logic: handle latency spikes and transient errors gracefully — retries, timeouts, degraded modes.
• Instrument cost: log timestamp + model + token count per request so you can attribute spend to peak vs off-peak windows the day a surcharge lands.

The Bottom Line

The July 24 alias retirement is the part that actually happened: every production integration now has to make an explicit Flash-versus-Pro choice. The surge pricing story is still a plan, not an invoice line. Treat it as something to be instrumented for, not something to reorganise your scheduling around — and check our pricing page, which tracks the official rate card weekly, before assuming otherwise.