Blog / · 6 min read
DeepSeek-V4 lands mid-July: 1M context, prices in freefall, and I'm still keeping a foot on the brake
1.6 trillion parameters, a context window announced at 1M tokens, prices that make your head spin. Except open-weight doesn't mean open-bar, and an announced context isn't a context you can actually use.
It’s July 12. DeepSeek is supposed to launch V4 “mid-July.” So right now, we’re talking about a model that hasn’t shipped in its final version yet.
There’s been a preview since late April, sure. But the official release, the one that will replace the old deepseek-chat and deepseek-reasoner endpoints, is a few days out.
I’m flagging this right away because on social media, people are already talking about V4 as if it’s a done deal, tested, benchmarked, deployed in prod at half the tech companies out there. No. This is an announcement.
Fine. Let’s still look at what’s inside, because the numbers are heavy.
What’s confirmed, and what isn’t yet
Two variants.
V4-Pro: 1.6 trillion parameters total, roughly 49 billion active per forward pass (it’s a MoE architecture, mixture of experts, so only a fraction of the model works on each token). V4-Flash: 284 billion total, roughly 13 billion active.
Both announce a 1 million token context window.
That part is confirmed by the official DeepSeek docs. Not a number floating around some affiliate blog.
What’s less confirmed: the final price. April’s preview ran on “classic” input/output pricing, cache hit and cache miss. But DeepSeek announced, for the mid-July launch, a dynamic pricing system: full rate between 9am-12pm and 2pm-6pm (Beijing time, I’d guess, to be confirmed once the docs are official), and half price outside those windows. The exact numbers for this new grid weren’t published yet at the time I’m writing this.
In other words: the “slashed price” everyone’s talking about is the preview price. The real price, the one that will run in prod, we don’t fully know yet.
I think this matters because it’s exactly the kind of detail that gets lost in “DeepSeek V4 is here, it’s ten times cheaper than GPT” articles. Maybe. But we don’t have the final bill yet.
What a 1M context actually means
Okay, 1 million tokens sounds like a dream. You dump an entire repo, your docs, your PR history, and the model is supposed to keep it all in mind.
The thing is, the announced size of a context window and the size you can actually use are two different things. It’s a known problem across the whole LLM industry, not specific to DeepSeek: the bigger the context gets, the more the model tends to lose precision on what’s buried in the middle, the more sensitive it becomes to the order in which you feed it information. This is called “lost in the middle” degradation in the literature. It’s not a conspiracy theory, it’s been documented across several model families for a while now.
I haven’t tested it on V4, the model isn’t even out in its final version. So I’m not going to tell you “V4 loses 30% accuracy at 500K tokens,” that would be made up. What I can say is that an announced 1M context is never automatically a usable 1M context. We’ll need independent evals, not just the vendor’s docs, to know where the tipping point sits on this specific model.
In the meantime, the sane instinct stays the same as before: give it targeted context rather than massive context. Dumping a whole monorepo “because you can” doesn’t necessarily improve the diff you get back. Mostly it inflates your bill and your review time, because you now have to check the model didn’t miss something buried at line 340,000.
Open-weight is not open-bar
Second point that bugs me a bit.
“Open-weight” gets waved around as an argument for total freedom. In practice, a 1.6T-parameter model, even MoE with 49B active, is still a beast to host. You’re not running this on your dev machine on a Friday night. You need a serious GPU cluster, or you go through a cloud host who, in turn, bills you, manages your quotas, and effectively becomes your new trusted intermediary.
Result: for the vast majority of teams, “open-weight” won’t change anything concrete. They’ll keep hitting the official DeepSeek API, or a cloud reseller’s. Which means: the same governance questions as with any closed model. Where does your data go. What retention policy. Which country hosts what. The fact that the weights are published answers none of these questions.
I already touched on this with Kimi K2.7 landing in GitHub Copilot: the real question is never “is the model open,” it’s “who decides which model runs on which repo, with what data, and who reviews the output.” Open-weight helps on one real point: you can, in theory, audit the weights, fork them, self-host if you have the means. But it’s not a blank check of trust. It’s still a model trained by a company, with its own choices, its own biases, and limits we still don’t fully know until there’s been a serious external evaluation.
What I’m taking away for an actual workflow
Honestly, the raw numbers are impressive. Nobody’s going to argue otherwise. But a parameter count doesn’t code in your place, and a 1M context doesn’t replace a human review.
What I’d do if I were on a team following this closely:
- Wait for the actual launch mid-July and the final pricing grid before running any ROI math. The “peak/off-peak” pricing changes everything if your usage is concentrated during office hours.
- Don’t migrate a critical pipeline onto it on day one. Give it two or three weeks for external evals to come out, beyond the marketing numbers.
- Keep treating long context as a tool to use sparingly, not as an excuse to stop breaking work into smaller chunks.
- Remember that “open-weight” doesn’t exempt you from asking the same governance questions you’d ask any vendor.
The “slashed cost” story deserves its own check once the real prices are published. I’ll run a concrete calculation at that point, the way I tried to do with the misleading reading of coding agent benchmarks: the tables are impressive, but the real test is your codebase, not the leaderboard.
For now, DeepSeek-V4 is a solid promise on paper. We’ll see in a few days what it’s worth once it’s plugged into real code.
Sources
Keep reading