The whole internet has been waiting for nearly three months!
DeepSeek V4 official version may be released as early as tomorrow, at the latest within the next few days.
Currently, some people have already received early access to the DeepSeek V4 (GA) grayscale testing.
There are two versions: DeepSeek V4 Flash and DeepSeek V4 Pro.
DeepSeek V4 'Full-Blooded Edition'
Is Finally Coming
What everyone is most concerned about is, how do you know if you've been included in the grayscale test?
Blogger AiBattle gave a folk 'verification mantra': look at the first person in the Chain-of-Thought (CoT).
If the model's thought process starts with "I'm" or "I'll", not the old version's "Let me".
Then congratulations, you're probably already using the V4 GA version!
In the current landscape where Fable 5 and GPT-5.6 Sol are locked in a fierce battle, the entire internet's expectations for the open-source AI's big move are sky-high.
A developer, Pankaj Kumar, after experiencing it, first gave a fair summary of V4's performance:
Overall performance approaches Opus 4.8 level, coding rivals GPT-5.6 Sol;
Agent capabilities are greatly enhanced, 3D and SVG generation have significantly improved;
For the same task, V4 requires more iteration rounds than Fable 5.
He stated that from the current market positioning, V4 will likely not beat the newly released Kimi K3, but the price will be significantly lower.
If this performance truly comes with this price, then it might be another DeepSeek moment.
First Round of Tests Already Out
Now, the first round of test demos for DeepSeek V4 (GA) have begun to circulate.
It can be said that the outside world has mixed reviews:
Some believe it can already rival Claude 5, but some developers also pointed out that the Pro version is not significantly ahead of the Flash version.
Below is a 3D simulated shooting game generated by V4 Pro. The core gameplay involves operating a ballista for target practice, and basic UI functions are already complete.
An HTML hybrid game blending Minecraft and No Man's Sky, created by the V4 official version, has relatively high playability.
The classic "Cut the Rope" game below was also produced by V4 in one go.
Below are also some demos generated by V4: an Xbox controller SVG test and game generation.
Introducing "Peak-Valley Billing" for the First Time
Still Smells Sweet
To recreate a "DeepSeek moment," the biggest variable is price.
All leaks point to the same conclusion: performance might not be number one, but the price will be significantly lower.
At the end of last month, DeepSeek sent an email to all API users, stating that the V4 official version would be launched in mid-July.
After the GA release, API pricing will be adjusted simultaneously, introducing "Peak-Valley Billing."
deepseek-v4-pro: $0.87 per million output tokens during off-peak, $1.74 during peak; $0.435 per million for cache-missed inputs during off-peak.
deepseek-v4-flash is even more aggressive: $0.28 per million outputs during off-peak, $0.56 during peak, $0.0028 per million for cache-hit inputs.
This means the "price butcher" who never raised prices has, for the first time, installed a meter on its computing power.
The impact on individual users is minimal, but for teams that run Agents non-stop during working hours, this is a real cost that needs to be recalculated.
Fortunately, the price for cache hits remains extremely low. Moving batch tasks, evaluation benchmarks, and data generation outside peak hours can still bring costs back down.
However, compared to Fable 5's $50 per million output tokens, V4 remains the most cost-effective option.
After all, the performance has also reached the Opus level. In the official self-test on SWE-bench Verified when the V4 preview was released—
DeepSeek-V4-Pro-Max was only 0.2 percentage points behind Claude Opus 4.6 Max, yet the price was only one-seventh of the latter's.
But in long-context retrieval, professional software engineering, and some knowledge reasoning evaluations, Opus still maintains a lead.
As a side note, the two old model names, deepseek-chat and deepseek-reasoner, will be officially discontinued on July 24th.
Regardless, this "shoe" that the whole internet has been waiting three months for is finally about to drop.
Judging from the currently leaked information, V4 will likely not be the model that is "number one in all categories."
With Fable 5 and GPT-5.6 Sol ahead, and Kimi K3 on the side, it's difficult for V4 to overturn the table relying solely on pure performance.
But DeepSeek has never played that card.
The real highlight is that old, repeatedly proven path: offering Opus-level capabilities at one-seventh the price.
After all, in the face of giants charging dozens of dollars per million tokens, even with a "peak-valley meter" installed, V4 is still that "price butcher" that wreaks havoc.
References:
https://x.com/pankajkumar_dev/status/2078536231026372846
This article is from the WeChat public account "New Zhiyuan," edited by Peach.
