The entire network has been waiting for nearly three months!
DeepSeek Version 4 official release may be launched as early as tomorrow, but no later than these days.
Currently, some people have already received theirs in advance. DeepSeek V4 (GA) grayscale testing access.
There are two versions in total: DeepSeek V4 Flash, DeepSeek V4 Pro.
DeepSeek V4 "Full-Power Version"
It's finally coming
What most people are probably most concerned about is how to know if they’ve been graylisted.
The blogger AiBattle provided a folk "inspection mantra": look at the first-person perspective in the chain of thought (CoT).
If the model's thought process begins with "I'm" or "I'll" instead of the previous version's "Let me."
Congratulations! You’ve likely already upgraded to the V4 GA version!
As Fable 5 and GPT-5.6 Sol continue to compete, the entire internet is brimming with anticipation for the big move in open-source AI.
After experiencing it, developer Pankaj Kumar provided the first balanced summary of V4's performance:
Overall performance approaches Opus 4.8 level, with encoding rivaling GPT-5.6 Sol;
Agent capabilities have been significantly enhanced, with 3D and SVG generation greatly improved;
For the same task, V4 requires more iteration rounds than Fable 5.
He stated that, based on the current market positioning, V4 is unlikely to outperform the newly released version. Kimi K3, but at a significantly lower price.
If this performance truly matches this price, it could be another DeepSeek Time.
The first round of testing is complete.
Today, DeepSeek The first-round test demo of V4 (GA) has begun to leak.
Opinions on this vary widely:
Some believe it has already matched Claude 5, but some developers also note that the Pro version does not significantly outperform the Flash version.
Below is a 3D simulation shooting game generated by V4 Pro, with the core gameplay centered on operating a siege crossbow for target practice; all basic UI functions have also been fully implemented.
The official V4 release features an HTML game that blends Minecraft and No Man's Sky, offering a high level of playability.
This classic game "Cut the Rope" was also generated in a single run by V4.
Below are some additional demos generated by V4: an Xbox controller SVG test and game generation.
Introducing "Peak-Valley Pricing" for the first time
Still just as good
Recreate DeepSeek The time" — the biggest variable is price.
All leaks point to the same claim: performance may not be the top priority, but the price will be significantly lower.
At the end of last month, DeepSeek An email was sent to all API users announcing that the official V4 release will go live in mid-July.
After the GA release, synchronize API pricing adjustments and introduce "peak and off-peak pricing."
DeepSeek-V4-Pro: $0.87 per million output tokens during normal hours, $1.74 during peak hours; $0.435 per million input tokens for cache misses.
DeepSeek-V4-Flash is even more aggressive: $0.28 per million output tokens, $0.56 during peak times, and $0.0028 per input token with cache hits.
In other words, the so-called "price butcher," who never raised prices, has for the first time installed a pricing meter on its computing power.
Has little impact on individual users, but for teams running Agents continuously during work hours, this is a real cost that needs to be recalculated.
Fortunately, the cached prices remain extremely low; moving batch tasks, benchmarking, and data generation outside peak hours can still bring costs back down.
However, compared to Fable’s 5 million output priced at $50, V4 still offers the best value for money.
After all, the performance has reached Opus level. When the V4 preview was released, the official self-test on SWE-bench Verified—
DeepSeek V4-Pro-Max is only 0.2 percentage points behind Claude Opus 4.6 Max, yet costs just one-seventh of the price.
However, Opus still leads in long-context retrieval, professional software engineering, and certain knowledge reasoning evaluations.
By the way, the two older model names, DeepSeek-Chat and DeepSeek-Reasoner, will be officially retired on July 24.
In any case, the much-anticipated "shoe" that the entire network has been waiting for three months is finally about to drop.
Based on the information currently available, V4 is unlikely to be the "all-around top" model.
Fable 5 and GPT-5.6 Sol lead the way, Kimi K3 is on the side, and V4 wants to shake things up with pure performance—it’s no easy feat.
But DeepSeek This was never the card we were playing.
The real highlight is the well-tested path: reducing the price to one-seventh while maintaining Opus-level capabilities.
After all, compared to giants with millions of tokens priced at dozens of dollars, even with a "time-of-use pricing meter" installed, V4 remains the dominant "price killer."
Reference materials:
https://x.com/pankajkumar_dev/status/2078536231026372846
This article is from the WeChat official account "NeurIPS", edited by Peach.