The next thing to watch from DeepSeek is not R2 or V5. It is the official release of V4 itself. Multiple outlets report DeepSeek is targeting a mid-July 2026 launch for the official version, and unlike a cosmetic "GA" stamp, it brings two changes you can budget around: peak-hour API pricing and stronger agentic, math, and code performance. What you can call today, deepseek-v4-pro and deepseek-v4-flash, is still the April 24 preview, and as of July 20, 2026 the official build has not appeared in DeepSeek's change log. This piece looks at what the official version changes, when it lands, and whether to build now or wait.
What the official DeepSeek V4 will change
Two things separate the coming official release from the preview that has been live since April. Both are concrete, not marketing.
Peak-hour pricing is the headline. In a move still rare for a frontier LLM API, DeepSeek is set to meter V4 by time of day. According to InfoWorld and Let's Data Science, calls during peak Beijing business hours (09:00–12:00 and 14:00–18:00) will cost twice the off-peak rate, while off-peak prices hold where they are today. The reported figures put V4-Pro output at ¥6 per million tokens off-peak versus ¥12 at peak, and V4-Flash at ¥2 versus ¥4. DeepSeek frames it as load-spreading for stability rather than a flat hike, but the practical effect is the same: interactive daytime traffic gets more expensive, and latency-tolerant or batch jobs that run off-peak stay cheap. Time-of-day pricing is normal in electricity and cloud-spot markets; seeing it on an LLM API signals that inference demand, not training, is now DeepSeek's binding constraint.
Performance goes up. The same reporting says the official build extends the preview with stronger agentic, math, and code-generation ability, and makes the 1M-token context window standard across the lineup. Treat the exact magnitude as unverified until independent benchmarks land, but the direction is a capability bump on top of the same two model IDs, not a new family.
What is *not* changing: the models stay open-weight, and you still call deepseek-v4-pro or deepseek-v4-flash. So the official release is a pricing-and-quality event, not a migration event.
When is the official version coming?
The short answer: reported for a mid-July 2026 window, but not yet confirmed shipped. As of July 20, 2026, that window is at or just past, and on the official change log the newest dated entry is still April 24, 2026, the preview. So the "V4 launches in July" chatter is half right: an official build is reported, but it is not live in the docs yet.
Two reliable signals will tell you the official version is genuinely live, rather than another rumor: a new dated change-log entry, and updated numbers on the official pricing page. When those move, it shipped. Until then, you are on the preview.
What V4 costs today, and after the official launch
Pricing is where the official release bites, so it is worth understanding the current baseline first. A lot of write-ups still quote DeepSeek-V4-Pro at $3.48 per million output tokens. That was the launch price and is long gone.
Here is the current pricing on DeepSeek's official pricing page, verified July 20, 2026 (all per 1M tokens, USD):
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 |
The reason Pro sits at $0.87 and not $3.48: a 75% launch discount cut it to one quarter of its original level, and when the promo ended on May 31, 2026, DeepSeek kept that lower rate as the new standard instead of letting it revert.
Now layer the official release on top. Once peak-hour pricing lands, today's $0.87 becomes the off-peak floor for Pro, and business-hours calls roughly double. For a chatbot serving users 9-to-5, that is a real increase; for an overnight batch pipeline, it changes almost nothing. Two levers keep spend flat: schedule tolerant work off-peak, and lean on cache-hit input pricing, which is roughly 50x cheaper than a cache miss on Pro. If you are modeling a real workload, the enterprise cost dynamics of running DeepSeek turn more on cache behavior, peak windows, and throughput limits than on the sticker price.
The July 24 deadline that is already fixed
One dated change is not a rumor: on July 24, 2026, 15:59 UTC, the legacy model names deepseek-chat and deepseek-reasoner retire. If your code still calls them, it breaks after that.
The migration is small, and doing it now also puts you on the model IDs the official release will use:
- Change the
modelparameter todeepseek-v4-flashordeepseek-v4-pro. - Keep the same
base_url(https://api.deepseek.com); it does not change. - Both models work through the OpenAI ChatCompletions interface and the Anthropic interface, so your client library stays put.
During the transition the old names map to the non-thinking and thinking modes of deepseek-v4-flash, so you are already effectively on V4 — the aliases just go away. If you route DeepSeek through an aggregator, the model-name change is the same; platforms like AIReiter expose the V4 IDs through the same Anthropic-compatible endpoint.
Build on the preview now, or wait for the official version?
For most teams, the answer is build now — with the official release priced in.
The preview is stable enough for production and, at $0.87 per million output tokens off-peak, cheap. V4-Pro at its maximum reasoning setting (a mode DeepSeek labels Pro-Max) has been reported around 80.6% on SWE-bench Verified, close to the strongest closed models; those figures are self-reported or third-party, not our own tests, so weight them accordingly. On cost, one Reddit thread summed up the pull as "17x cheaper" than what people were paying elsewhere.
Waiting for the official build buys you a performance bump but also the peak-hour surcharge, and it delays work for a launch that, as of July 20, has not yet posted to the change log. The pragmatic path: ship on the preview, budget for peak pricing on interactive traffic, and re-test when the official numbers post.
Which model? Developers on Reddit describe Flash as a "super cheap worker" for well-scoped tasks and reach for Pro on frontier-grade reasoning; the Flash-versus-Pro trade-offs come down to that split. Two caveats hold either way: V4-Pro throughput is capacity-constrained (DeepSeek has said as much), so Pro can be slower under load, and V4 is text-only — image, audio, and video live in separate DeepSeek lines (Janus, DeepSeek-VL2, DeepSeek-OCR), and the official release is not expected to change that.
What about R2 and V5?
Neither is the near-term story, despite the search volume. R2 is unreleased with no confirmed date; DeepSeek has denied specific windows, and the leaked "specs" contradict each other (a ~1.2T MoE model in one leak, a 32B dense model in another). V5 is thinner still: the viral "V5 on July 24" claim is a misreading of the legacy-alias retirement date. The credible next milestone is the one this article is about — the official V4 — not a new generation.
FAQ
When is the official DeepSeek V4 coming out?
Multiple outlets reported a mid-July 2026 window, but it had not appeared on the API change log as of July 20, 2026. Watch the change log and the pricing page for confirmation that it has actually shipped.
What's new in the official V4 versus the preview?
Two things: peak-hour API pricing (business-hours calls roughly double the off-peak rate) and stronger agentic, math, and code performance, with 1M context standard across the lineup. The model IDs and open-weight licensing stay the same.
How much will the official DeepSeek V4 cost?
Off-peak pricing is expected to match today's rate: $0.87 per 1M output tokens for V4-Pro, $0.28 for V4-Flash. Peak-hour calls are set to cost about twice that. Cache-hit input is far cheaper on both.
Is DeepSeek V4 usable right now?
Yes. V4-Pro and V4-Flash have been public, open-weight (MIT-licensed, on DeepSeek's Hugging Face org), and callable since April 24, 2026. You do not have to wait for the official release to build on them.
Do I need to migrate before July 24, 2026?
Only if your code still calls deepseek-chat or deepseek-reasoner. Switch the model name to deepseek-v4-flash or deepseek-v4-pro before July 24, 2026, 15:59 UTC; nothing else in your request changes.