DeepSeek V4 Ships With Time-of-Day Pricing — And a July 24 Deadline That Breaks Old Code

DeepSeek has moved its V4 model line out of preview and into a full mid-July 2026 launch, and it arrives with two changes developers actually have to plan around: a pricing scheme that charges twice as much during business hours, and a hard cutoff that retires the model names a lot of existing code still uses. Both are confirmed on DeepSeek's own documentation. The performance claims around V4 are noisier, and this article keeps those separate from the facts you can bank on.

Two tiers, and what they actually are

DeepSeek's official API docs list two production models. deepseek-v4-pro is the large tier at 1.6 trillion total parameters with 49 billion active per token, and deepseek-v4-flash is the lightweight tier at 284 billion total / 13 billion active (DeepSeek API docs). "Active parameters" is the part that runs for any single token in a mixture-of-experts model, which is why Flash is cheap to serve despite still being a large model on paper.

The base, off-peak prices come straight from DeepSeek's official pricing page, quoted in US dollars per one million tokens (DeepSeek pricing):

Those numbers are worth internalizing for one reason: the gap between a cache hit and a cache miss is roughly 50x on input. DeepSeek rewards you heavily for reusing prompt prefixes it has already seen, so how you structure repeated requests matters more to your bill than which tier you pick.

The new twist: peak-hour pricing

The headline change is that V4 introduces peak and off-peak API pricing — the first time DeepSeek has charged by time of day. During peak hours, API usage costs twice the off-peak rate. Peak windows run from 9:00 a.m. to 12:00 p.m. and 2:00 p.m. to 6:00 p.m. Beijing time, and everything outside those windows bills at the standard rate above (TechNode, corroborated by KuCoin).

The practical read: DeepSeek is doing demand-shaping. If your workload is batch — evaluations, data labeling, offline generation — scheduling it outside Beijing business hours roughly halves the cost. If your workload is interactive and user-facing, you can't dodge peak, and you should budget for the higher number. This also turns time-of-day into a routing dimension for anyone running a multi-model gateway (TheRouter.ai).

The deadline that will break code: July 24

Here is the part that is easy to miss and expensive to ignore. When DeepSeek shipped V4, it kept the old model aliases — deepseek-chat and deepseek-reasoner — working as a courtesy, transparently routing them to V4-Flash during a grace period. That grace period ends. Per DeepSeek's official notice, both legacy names are fully retired and inaccessible after July 24, 2026, 15:59 UTC (DeepSeek API docs). After that timestamp, any request still using those names fails outright.

The fix is small but non-optional: swap deepseek-chat for deepseek-v4-flash, deepseek-reasoner for the thinking variant, and use deepseek-v4-pro where you want the larger model (Enterprise DNA). Same base URL, same API key, same request structure — only the model string changes. If you or a vendor you depend on hard-coded the old names, that one string is the difference between working and broken this week.

On performance — read this part skeptically

Secondary coverage repeats that V4 Pro scores 80%+ on SWE-bench, the software-engineering benchmark, and some write-ups frame it as closing on frontier Western models. Treat that as vendor-reported and not independently verified: DeepSeek's own release notes present benchmark results as images rather than auditable numbers in text (DeepSeek API docs), and no outside lab has published a reproduction as of this writing. Any claim that V4 reaches parity with a specific model like Claude Opus is, right now, a vendor-side comparison with no third-party benchmark behind it — worth noting, not worth betting on.

What this means for vybecoding users

Nothing on vybecoding.ai routes to DeepSeek today, so there is no action for readers. The reason we're flagging V4 is the pattern, not the model: time-of-day pricing is a new lever that will show up in more providers, and "the old model name stopped working" is the kind of silent breakage that takes down pipelines quietly. If you build with third-party AI APIs on your own projects, the takeaways are (1) check whether anything you run still calls a deprecated model name before July 24, and (2) if you do batch work against DeepSeek, schedule it off-peak. We'll revisit V4 only if independent SWE-bench numbers land — until then, the pricing and the deadline are the real story.