DeepSeek V4 goes fully stable, and the old models switch off today

DeepSeek’s V4 model line has now moved from preview to a full production release, and as part of that switch the company’s older chat models are being retired today, July 24. If you or a tool you use still points at the legacy “deepseek-chat” or “deepseek-reasoner” names, those stopped working at 15:59 UTC. The fix is simply to move to V4, which has been the recommended default for a while.

The details are worth a look because DeepSeek keeps doing the thing that shakes up the market: strong models at prices that make competitors flinch. V4 comes in two sizes. V4-Pro is a large mixture-of-experts model (a design where only a slice of the network, here about 49 billion of its 1.6 trillion parameters, activates for any given request, which keeps it fast and cheap to run) priced around $0.44 per million input tokens and $0.87 per million output tokens. V4-Flash is the smaller, cheaper tier at roughly $0.14 and $0.28. Both handle a 1M-token context window (very roughly, how much text the model can hold in mind at once), and both ship as open weights under the permissive MIT license, meaning anyone can download and run them.

So what is really going on here? A “stable” release sounds like a technicality, but it clears the last hesitation for companies that were waiting for a non-preview label before putting DeepSeek into real products. And the pricing is the quiet headline. When a capable open model costs well under a dollar per million output tokens, every other lab gets measured against that number. It is a big reason output prices across the industry have been sliding all year. Worth keeping expectations grounded, though: cheapest does not mean best at everything, and independent benchmarks still put the top frontier models ahead on the hardest reasoning and coding tasks.

What this means for you: If you use an app built on DeepSeek, you might see a brief hiccup today if the developer was slow to update, but nothing you need to do yourself. For the curious, this is a good example of why open-weight models matter: because anyone can run them, they push prices down for everyone, including the paid tools you already use. For power users and small teams, V4 is one of the cheapest ways to get frontier-adjacent quality, and running it yourself keeps your data on your own hardware. For most people, the takeaway is simpler: AI is getting cheaper fast, and that is mostly good news.

Sources

• DeepSeek API documentation and changelog
• OpenRouter: DeepSeek V4 Pro pricing and benchmarks