New Chinese AI Model Beats OpenAI, Google and Microsoft by a Mile With 214x Cheaper Pricing, 100% Faster and 14.8 Trillion Training Tokens

Artificial intelligence (AI) is quickly becoming one of the most powerful forces in finance, with trillion-dollar companies like Apple (AAPL), Google (GOOGL), and Microsoft (MSFT) all racing to take the lead in what many are calling the new “AI gold rush.” These advanced models are already being used to analyze markets, predict trends, and automate complex investment decisions faster than any human ever could. As the technology becomes cheaper and more accessible, everyday investors may soon have tools once reserved for elite hedge funds and billion-dollar giants. But with so much money and hype pouring into the space, it’s also creating a more unpredictable and fast-moving market environment.

In a major shift within the generative AI race, Chinese firm DeepSeek has released a new model, DeepSeek-V3, that decisively outperforms its U.S. counterparts — including OpenAI’s GPT-4 Turbo, Google’s Gemini 1.5, and Anthropic’s Claude 3 Opus — across key metrics such as benchmark scores, training scale, speed, and cost.

Benchmark Dominance

According to recent data, DeepSeek-V3 and its sibling models are scoring at the top of multiple evaluation datasets:

• MMLU (Massive Multitask Language Understanding): DeepSeek-V3 reaches 88.5%, compared to GPT-4’s 86.4% — a 2.1 percentage point lead on this general-knowledge benchmark.
• HumanEval (Code Generation): DeepSeek-V3 achieves 82.6% pass@1, outperforming GPT-4’s 67% by a significant 23% margin, indicating superior zero-shot coding capabilities.
• HellaSwag (Commonsense Reasoning): DeepSeek-R1 scores 95.3%, while GPT-4 generally performs in the 90% range, offering a 5.3% advantage.
• GSM8K (Grade School Math): DeepSeek-V3 reportedly scores 81.5%, surpassing comparable Gemini and Claude 3 Opus results, which typically hover in the mid-70s.

Speed and Efficiency

One of DeepSeek’s standout advantages is speed. DeepSeek-V3 delivers answers with a median latency of just 0.8 seconds, compared to 1.6–1.8 seconds for GPT-4 Turbo — offering a 100% faster response time in practical deployments.

This efficiency is enabled by a Mixture-of-Experts (MoE) architecture that activates only a fraction of the total model per query — typically 4 out of 64 experts — allowing for both scalability and resource optimization.

Scale and Data Volume

DeepSeek’s performance is underpinned by a vast training corpus. V3 was trained on approximately 14.8 trillion tokens, based on disclosures from company technical papers and corroborated in performance comparisons — a figure 6-12x larger than the 1–2 trillion tokens estimated for GPT-4.

Don't Miss: 

• Think it’s too late to invest in the booming AI sector? This one’s still under the radar

Additionally, DeepSeek models are capable of handling context windows exceeding 200,000 tokens, doubling the 128K-token limit found in GPT-4 Turbo.

Cost Leadership

Pricing is another key differentiator. DeepSeek’s base models offer input token rates as low as $0.14 per 1M tokens, versus $10.00 per 1M tokens for GPT-4 Turbo — a 93.6% cost advantage or 214x cheaper. 

Outlook

Given its pace of technical advancement, DeepSeek is now viewed as a serious challenger to the AI dominance historically held by U.S. firms.

The initial release of DeepSeek caused markets to drop, with Nvidia specifically seeing a substantial decline in its stock price. However, DeepSeek did not have a long-term impact on the AI sector, or investor enthusiasm for AI in general. 

As its models top leaderboards and its architecture proves scalable, DeepSeek’s rise signals a broader shift in the balance of AI power — from not just competition, but potential disruption of Western AI supremacy.