Skip to content
NewsProduct launch

DeepSeek V4 transition complete: legacy endpoints gone, announced peak-time surcharge has not yet activated

· by Pondero Newsdesk

The short version

DeepSeek's deepseek-chat and deepseek-reasoner API aliases were retired on July 24, 2026 at 15:59 UTC with no deadline extension. An announced time-of-day surcharge for Beijing business hours was not yet active as of the migration date, per the official rate card.

DeepSeek V4 transition complete: legacy endpoints gone, announced peak-time surcharge has not yet activated

DeepSeek retired its deepseek-chat and deepseek-reasoner API aliases at 15:59 UTC on July 24, 2026, with no extension to the deadline set in the official changelog since April. Any production integration still calling those model names is now returning errors.

What went live

The retirement ended three months of silent compatibility routing. Since V4's preview launch on April 24, 2026, deepseek-chat mapped to V4-Flash in non-thinking mode and deepseek-reasoner mapped to V4-Flash in thinking mode, per the pricing page footnote. That silent layer is gone.

Two endpoints remain active: deepseek-v4-pro and deepseek-v4-flash. V4-Pro carries 1.6 trillion total parameters with 49 billion active per forward pass in a mixture-of-experts configuration and a 1-million-token context window. V4-Flash runs 284 billion total with 13 billion active at the same 1M context length, targeting cost-sensitive and high-throughput workloads.

Current flat rates per the official rate card: V4-Pro at $0.87 per million output tokens and $0.435 per million input tokens on a cache miss; V4-Flash at $0.28 per million output and $0.14 per million input. Cache hits drop to $0.003625 and $0.0028 respectively.

What has not gone live: the announced peak-time surcharge

DeepSeek disclosed plans for time-of-day pricing around the V4 release. Per TechNode's June 30 reporting, the company announced that API usage during Beijing business hours (9 a.m. to noon and 2 p.m. to 6 p.m., Beijing time) would be billed at twice the off-peak rate. The current pricing page shows a single flat tier with no time-of-day component. DeepSeek has not published a specific activation date for the surcharge.

The mechanism would replace the prior off-peak discount program, which also ran flat and has since ended. An activated surcharge at those hours would be a first among major AI API providers for time-based demand pricing.

Why it matters for operators

The immediate check: is your integration currently failing? Teams that did not update model strings before the July 24 cutoff will see errors on every call to the old aliases. The repair is a one-line change. The base URL (https://api.deepseek.com), the API key, and all other request parameters are unchanged. Teams choosing between V4-Pro and V4-Flash should note that deepseek-reasoner callers were previously receiving V4-Flash behavior, not V4-Pro. Switching explicitly to deepseek-v4-pro will change output and raise per-call cost.

The forward consideration is batch scheduling. The announced peak surcharge has not gone live, so there is no live cost incentive today to shift workloads off Beijing business hours. The practical move now is instrumentation: logging timestamps alongside token counts per request means any surcharge that does activate can be attributed to peak versus off-peak windows without retrofitting.

What to watch next

Teams still on legacy aliases face a hard break: calls are failing now. The repair has two steps. First, choose a model: deepseek-v4-flash for any workload that ran on deepseek-chat or deepseek-reasoner at the same price tier ($0.28 per million output tokens), or deepseek-v4-pro for quality-sensitive, long-context, and agentic work where V4-Flash behavior no longer cuts it ($0.87 per million output tokens, three times the cost). Second, instrument: add a timestamp to every logged API call. The announced peak-hour surcharge has no activation date, but when it lands, the window to retrofit logging will be narrow.

Watch the DeepSeek pricing page for any time-of-day rate structure. If the surcharge activates, watch whether other inference providers adopt similar demand-pricing models.

Sources