DeepSeek V4-Pro went GA on August 13, 2026 — as of this post, two days ago. If you've been half-following the release, that sentence probably needs unpacking, because "V4 launched" doesn't map to a single date the way model releases used to.
Here's the actual timeline, as far as I can piece together from DeepSeek's own changelog and release notes:
April 24, 2026 — Both V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B total / 13B active) shipped as an open-weight preview, MIT licensed, both with 1M-token context.
July 31, 2026 — V4-Flash got an official, non-preview release (build "0731"), described as the same architecture and pricing as the preview, just re-post-trained with a focus on agentic capabilities.
August 13, 2026 — V4-Pro finally went GA (designated V4-Pro-0813), more than three and a half months after its preview, also with a stated focus on agent capabilities and published benchmarks on tasks like Terminal Bench and repo-level coding.
August 16, 2026 — a price increase for V4-Pro is scheduled to take effect, per DeepSeek's own announcement.
That's not one launch. It's four distinct events spread across four months, for what most people are casually calling "DeepSeek V4" as if it were a single release. I think that's worth noticing, because it's not unique to DeepSeek — it looks like the shape most major model releases are taking now, and it changes how developers should think about "using the latest model."
Releases used to be an event. Now they're closer to a pipeline.
The old mental model — a model gets announced, benchmarked, and made available, roughly all at once — still shapes a lot of how people talk about new releases ("DeepSeek V4 is out, is it better than X?"). But what actually happened here is closer to a staged rollout: preview first (to let the ecosystem start testing against open weights), a smaller/cheaper variant graduating to GA first, the flagship variant following separately, and a price change arriving after adoption has had time to build.
This isn't necessarily a bad pattern. Shipping the smaller Flash variant to GA first plausibly lets a provider validate real-world agentic performance and pricing at lower risk before committing the flagship to the same treatment. But it does mean that questions like "is DeepSeek V4 good?" or "what does it cost?" don't have a single stable answer for months after the first headline — the honest answer depends on which variant, and which point in its rollout, you're actually asking about.
What this means if you're building on it
A few practical implications I'd draw from this pattern, not specific to DeepSeek:
"Preview" and "GA" can behave differently even with an unchanged model name. DeepSeek's own notes are explicit that the GA build of Flash kept the same architecture and parameter count but was re-post-trained — meaning behavior on your specific prompts could shift between preview and GA even without a version-number change you'd notice unless you were watching the changelog.
Published benchmarks often lag the model you're actually calling. Benchmark numbers attached to a GA announcement typically reflect that specific build. If you tested against the April preview and are now calling the August GA build, your own results are the more reliable signal, not the announcement numbers.
Pricing tied to a rollout stage is a moving target. A price increase scheduled days after GA is a reminder that the number you see on launch day isn't necessarily the number you'll be paying against in production a few weeks later — worth checking current pricing before treating any specific figure as fixed in an architecture decision.
None of this is a criticism of DeepSeek specifically — I'd expect similar staggered patterns from most large model providers going forward, since it's a reasonable way to de-risk a big launch. But it does mean "I'm using the latest model" is a less stable claim than it used to be, and it's worth being specific about which build and which date you tested against when comparing notes with other developers.
For what it's worth, when I wanted to compare behavior across the Flash and Pro variants during their respective rollouts, I used a single OpenAI-compatible gateway (RouteAI, in my case) so I could point the same test harness at different model identifiers without maintaining separate client setups for each — useful mainly because the model names themselves changed a few times over the rollout, which is exactly the kind of churn a shared interface absorbs more easily than hardcoded per-provider clients.
The takeaway
"DeepSeek V4 is out" undersells what's actually a multi-stage rollout spread across months, with different variants, different GA dates, and pricing that's still moving as of this week. If you're evaluating or writing about DeepSeek V4 — or honestly, any major model release from here on — it's worth being explicit about which variant and which build date you mean, since the gap between "previewed" and "GA, stable, and priced" can now stretch for months.
Has anyone else noticed this staged-rollout pattern becoming more common across providers, or is DeepSeek's timeline here more staggered than most?
TL;DR: DeepSeek V4 wasn't a single launch — it was four separate events over four months (April preview, July Flash GA, August Pro GA, an August 16 price change). Major model releases are increasingly staged rollouts rather than single events, which means benchmarks, pricing, and even model behavior can shift between "preview" and "GA" — worth checking which specific build you're actually testing before drawing conclusions.
Here's the tool I referenced in this post: www.fastrouteai.com
Top comments (1)
If you’re reading this after publication and the pricing or model versions have changed, please refer to the official documentation for the latest information. I’ll update this post when there are major changes.