DeepSeeek V4 has been released, highlighting some groundbreaking victories against Gemini, ChatGPT, and Claude.
A new open-source model that surpasses GPT-5.4 on Codeforces and is available at a fraction of Claude's cost is not merely a news item; it's a significant challenge to the entire AI industry's pricing structure.
Nadeem Sarwar / Digital Trends
China’s DeepSeek frequently makes an unexpected entrance at Silicon Valley’s AI events, and this time, it has arrived with the much-anticipated V4 preview. The Hangzhou-based company has unveiled its latest AI model, which outperforms well-known American models in specific areas.
DeepSeek has introduced two new models: V4-Pro (Expert mode) and V4-Flash (Instant mode). The former is an extensive 1.6 trillion parameter model, while the latter has a more manageable 284 billion parameters. Both models boast a one-million-token context window.
🚀 DeepSeek-V4 Preview is officially released and open-sourced! Welcome to the age of cost-effective 1M context length.🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance on par with the leading closed-source models globally.🔹 DeepSeek-V4-Flash: 284B total / 13B active params.… pic.twitter.com/n1AgwMIymu— DeepSeek (@deepseek_ai) April 24, 2026
What exactly did DeepSeek unveil?
Of particular significance is that both models are open source, thus they can be downloaded from Hugging Face and operated locally on personal hardware. However, the substantial scale of V4-Pro necessitates a considerable amount of VRAM for local execution.
One of the most captivating aspects of the announcement is its comparison to popular AI models like Gemini, ChatGPT, and Claude. For instance, V4-Pro excels in coding, achieving a Codeforces rating of 3,206, surpassing GPT-5.4's score of 3,168 and Gemini 3.1's score of 3,052. This positions it as the most powerful open model for competitive programming activities.
On LiveCodeBench, V4-Pro achieves a score of 93.5, which puts it ahead of Claude Opus 4.6's 88.8 and Gemini's 91.7. Similarly, it scores 51.8 on Toolathlon for agentic tasks, outpacing both Claude (47.2) and Gemini (48.8). Meanwhile, the faster and more efficient V4-Flash matches V4-Pro on simpler agent tasks at a significantly lower compute cost.
In which areas does V4-Pro outperform its rivals?
Benchmark | DeepSeek V4-Pro | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro
--- | --- | --- | --- | ---
Codeforces (Rating) | 3,206 | — | 3,168 | 3,052
LiveCodeBench (Pass@1) | 93.5 | 88.8 | — | 91.7
Apex Shortlist (Pass@1) | 90.2 | 85.9 | 78.1 | 89.1
SWE Verified (Resolved) | 80.6 | 80.8 | — | 80.6
Toolathlon (Pass@1) | 51.8 | 47.2 | 54.6 | 48.8
Terminal Bench 2.0 (Acc) | 67.9 | 65.4 | 75.1 | 68.5
MRCR 1M Long Context | 83.5 | 92.9 | — | 76.3
HMMT 2026 Math | 95.2 | 96.2 | 97.7 | 94.7
IMOAnswerBench | 89.8 | 75.3 | 91.4 | 81.0
Nevertheless, there are several areas in which DeepSeek’s new model falls short of its competition. For example, Claude’s Opus 4.6 excels in long-context retrieval, scoring 92.9 on MRCR 1M compared to V4-Pro’s 83.5. GPT-5.4 continues to lead on Terminal Bench 2.0 with a score of 75.1 against V4-Pro’s 67.9.
Where DeepSeek truly revolutionizes the competition is in pricing. The V4-Pro model is priced at $3.48 per million output tokens, which, in contrast to OpenAI’s $30 and Anthropic’s $25 for similar workloads, is significantly more appealing to potential customers. This disparity is substantial for everyday developers creating AI-powered applications.
Other articles
One of the top gaming CPUs ever produced has just seen a price drop of $60: the AMD Ryzen 7 7800X3D is now available for $388.
Since its release, the AMD Ryzen 7 7800X3D has consistently ranked as the top gaming CPU, and currently priced at $388.98, it is $60 lower than its original list price of $449. For those constructing or enhancing an AM5 system with a focus on gaming performance, the Ryzen 7 7800X3D continues to be the ideal processor, and this discount makes the […]
You can now arrange your Spotify playlists into folders on your mobile device.
Spotify's updated mobile playlist folders allow users to organize playlists, play all items within a folder simultaneously, and even shuffle the entire collection.
The value of the US government's stake in Intel has reached $36 billion. This outcome was not part of any plans made in Washington.
The U.S. government’s 9.9% stake in Intel, obtained for $8.9 billion through converted CHIPS Act funds, has risen in value to $36 billion following a surge in Q1 earnings. There is currently no exit strategy in place.
The EU's efforts to enhance child safety have hit a roadblock as the ePrivacy derogation comes to an end, an age verification application has been compromised, and the CSA Regulation remains stalled in trilogue discussions.
The ePrivacy derogation ended on April 3, the EU age verification application was compromised in just minutes, and the CSA Regulation is still at a standstill. Europe's child protection measures rely on data that its privacy regulations prohibit.
Despite rumors of reductions, Apple may still deliver a performance showcase with the iPhone 18.
Apple's iPhone 18 could put an end to cost-cutting doubts with its 2nm A20 chip, 12GB of RAM in the standard versions, variable-aperture cameras in the Pro models, and Apple's proprietary C2 modem.
X-Energy secures $1.02 billion in a historic nuclear initial public offering, with the Amazon-supported reactor developer experiencing a 31% jump upon its debut on Nasdaq.
X-Energy's IPO of $1.02 billion saw demand that was 15 times greater than the shares available and was priced 21% higher than its initial range. This is notable given that the same company was unable to secure a SPAC deal at $1 billion in 2023. The demand for AI data centers has significantly altered the landscape.
DeepSeeek V4 has been released, highlighting some groundbreaking victories against Gemini, ChatGPT, and Claude.
DeepSeek V4-Pro achieves a score of 3,206 on Codeforces, surpassing GPT-5.4 and Gemini, and is priced at $3.48 per million tokens compared to Claude's $25, positioning it as one of the most competitively priced frontier-class AI launches of 2026.