DeepSeek V4 "full-power version" has been exposed and is expected to be released as soon as tomorrow.

The entire online community has been waiting for nearly three months!

The official release of DeepSeek V4 may happen as early as tomorrow, and no later than the next few days.

At present, a number of users have been granted early access to the gray test of DeepSeek V4 (GA).

There are two versions in total: DeepSeek V4 Flash and DeepSeek V4 Pro.

The Full-Power Version of DeepSeek V4

Is Finally Coming

What everyone is most concerned about must be: how to check if you have been selected for the gray test?

Blogger AiBattle shared a community-proven quick check trick: look at the first-person wording in the Chain of Thought (CoT).

If the model's thinking process starts with "I'm" or "I'll" instead of the old version's "Let me".

Then congratulations, you are very likely already using the V4 GA version!

Against the backdrop of ongoing competition between Fable 5 and GPT-5.6 Sol, the entire online community has extremely high expectations for this major new release in the open-source AI space.

After testing it, developer Pankaj Kumar first gave a fair summary of V4's performance:

The overall performance is close to the Opus 4.8 level, with coding capabilities rivaling GPT-5.6 Sol;

Agent capabilities have been greatly enhanced, and 3D and SVG generation performance has been significantly improved;

For the same task, V4 requires more iteration rounds than Fable 5.

He said that based on the current market positioning, V4 is most likely not as powerful as the newly released Kimi K3, but its price will be significantly lower.

If this level of performance is really paired with this price, it could mark another landmark DeepSeek moment.

First Round of Test Results Released

Now, the first round of test demos for DeepSeek V4 (GA) has started to circulate publicly.

Public opinions on it are mixed:

Some people believe its performance can already match Claude 5, but some developers also point out that the Pro version does not have a significant advantage over the Flash version.

The following is a 3D simulated shooting game generated by V4 Pro. Its core gameplay is to control a siege crossbow vehicle to complete target practice, and all basic UI functions are fully implemented.

The official V4 release also created a playable HTML game that combines the mechanics of Minecraft and No Man's Sky, which has a fairly high level of playability.

The following classic game Cut the Rope was also generated in a single run by V4.

The following are more demos generated by V4, including an SVG test of an Xbox controller and another game generation result.

Peak-Valley Pricing Introduced for the First Time

Still Extremely Cost-Effective

The biggest variable that could create another landmark DeepSeek moment is its pricing.

All leaked information points to the same conclusion: its performance may not rank first, but its price will be significantly lower than competitors.

At the end of last month, DeepSeek sent an email to all API users, stating that the official V4 version will be launched in mid-July.

After the GA release, API pricing will be adjusted simultaneously, with the introduction of peak-valley billing.

For deepseek-v4-pro: the price for 1 million output tokens is $0.87 during off-peak hours and $1.74 during peak hours; the price for cache-missed input is $0.435 per million tokens during off-peak hours.

The pricing for deepseek-v4-flash is even more aggressive: $0.28 per million output tokens, $0.56 during peak hours, and only $0.0028 per million cache-hit input tokens.

In other words, the "price slasher" that has never raised prices before has installed a metering mechanism for its computing power for the first time.

This change has little impact on individual users, but for teams that run Agent tasks continuously during working hours, this is a real cost that needs to be recalculated carefully.

Fortunately, the price for cache hits remains extremely low. If you move batch tasks, benchmark testing, and data generation to non-peak hours, you can still keep costs under control.

However, compared to Fable 5's pricing of $50 per million output tokens, V4 still has the most outstanding cost-effectiveness.

After all, its performance has reached the Opus level. When the preview version of V4 was released, in the official self-tested SWE-bench Verified benchmark —

DeepSeek-V4-Pro-Max was only 0.2 percentage points behind Claude Opus 4.6 Max, but its price was only one-seventh of the latter.

However, Opus still maintains a leading position in long-context retrieval, professional software engineering, and some knowledge reasoning benchmarks.

By the way, the two old model names deepseek-chat and deepseek-reasoner will be officially taken offline on July 24.

In any case, the long-awaited new release that the entire online community has waited three months for is finally about to arrive.

Based on the currently circulated information, V4 is most likely not going to be the model that ranks first in all indicators.

With Fable 5 and GPT-5.6 Sol as direct competitors, and Kimi K3 as a strong rival, it is very difficult for V4 to stand out solely by pure performance.

But DeepSeek has never relied on this strategy.

The real highlight is its repeatedly proven mature approach: delivering Opus-level capabilities at one-seventh of the price.

After all, in a market where major players often charge dozens of dollars per million tokens, even with the new peak-valley pricing mechanism, V4 is still the dominant "price slasher" in the industry.

References:

https://x.com/pankajkumar_dev/status/2078536231026372846

This article is from the WeChat public account AI Era, edited by Taozi, and published with authorization from 36Kr.