DeepSeek Strikes Back: China’s Ambitious New AI Models Aim Directly at Gemini and ChatGPT

For most of 2025, the global AI narrative has felt predictable. Google’s Gemini family kept stretching its lead with increasingly powerful text-to-image and reasoning capabilities, while OpenAI continued its agentic push with GPT-5 and the recently upgraded GPT-5.1. Even China’s fast-growing AI landscape looked settled, with players like Qwen and Kimi seizing attention with their impressive leaps in image, text and multimodal performance.

And then, just when the industry thought the hype around DeepSeek had peaked earlier this year, the Chinese startup has returned — loudly, ambitiously, and with a renewed claim that could shake the foundations of the East-vs-West AI race.

DeepSeek is back. And this time, the company says it’s bringing GPT-5-level power to the table.

DeepSeek: A comeback moment after months of silence

DeepSeek burst onto the global AI map in early 2025 when its R1 and V3 models stunned analysts by matching Western-built models — at a fraction of the cost. For a brief moment, it felt as if the centre of gravity in the AI world had shifted.

But the momentum didn’t last long.

As Google pushed out its Gemini 3 Pro and its Nano line, and OpenAI doubled down on its agentic workflows, DeepSeek seemed to disappear from the battlefield. Meanwhile, Chinese rivals Qwen and Kimi began leading the domestic narrative with strong text and image model families.

Industry watchers wondered: Was DeepSeek a one-season phenomenon?

Today, the startup has delivered a very clear answer.

The big reveal: DeepSeek V3.2 and V3.2 Speciale

DeepSeek has launched two new AI models — V3.2 and V3.2 Speciale — designed explicitly to take on Google’s Gemini line and OpenAI’s GPT-5 family.

The company describes them as reasoning-first models built primarily for agentic workflows — a segment where global giants have been aggressively innovating.

Performance claims that demand attention

According to DeepSeek:

• V3.2 delivers general-task performance comparable to GPT-5.

• V3.2 Speciale matches the reasoning capabilities of Google’s Gemini 3 Pro, particularly in complex, multi-step problem-solving.

If true, this places DeepSeek back in the top tier of global AI capabilities — a position it briefly held earlier this year.

But the most headline-grabbing achievement is something else entirely.

A breakthrough in elite math benchmarks

DeepSeek says that the V3.2 Speciale model secured gold-medal-level performance in:

• 2025 International Mathematical Olympiad (IMO)

• 2025 International Olympiad in Informatics (IOI)

Until now, such accomplishments were exclusive to proprietary, frontier-scale models from Google and OpenAI. For DeepSeek to break into this space signals a serious escalation in China’s ambitions for mathematically rigorous AI systems.

The company also claims that V3.2 Speciale:

• Outperforms GPT-5 High and Gemini 3 Pro on pure math benchmarks like AIME and HMMT

• Leads on pure coding logic benchmarks such as CodeForces, surpassing models from every major AI lab

This places DeepSeek’s new release among the world’s strongest publicly documented math-reasoning AIs.

But it’s not a clean sweep

On the highly challenging Humanity’s Last Exam benchmark — a frontier test gaining attention this year — DeepSeek V3.2 Speciale managed to score higher than GPT-5 High, but still fell short of Gemini 3 Pro.

This suggests that the global race is tighter than ever, with each model dominating its own niche.

How users can access the new models of Deepseek

DeepSeek is rolling out access in phases:

• DeepSeek V3.2
  Now live as the default model on the DeepSeek website and app, replacing the earlier V3.2 Experimental model introduced in September.

• DeepSeek V3.2 Speciale
  Available only through API for now.
  There is no confirmed timeline for public access through the chatbot interface.

For everyday users, this means the high-performance Speciale variant is still in a limited release — a move that mirrors what Google and OpenAI often do with their early-stage, high-power variants.

The return of DeepSeek marks an important moment in the global AI landscape:

1. It reopens a competitive front in reasoning-first agentic AI.
   After months of relative stagnation, DeepSeek is signalling that it is not done challenging American and Chinese rivals.

2. It strengthens China’s position in math-intensive AI research.
   Achieving Olympiad-level performance with a startup budget is a powerful narrative.

3. It pressures OpenAI and Google to continue innovating.
   With the lines blurring between proprietary labs and cost-efficient challengers, competitive dynamics are shifting rapidly.

4. It reintroduces DeepSeek as a serious player, after speculation that the startup had fallen behind.

The outlook: A crowded battlefield gets even tighter

2025 has already been a historic year in the AI race, defined by rapid model cycles, agentic workflows, reasoning benchmarks and cost-efficiency breakthroughs.

With V3.2 and V3.2 Speciale, DeepSeek is signaling a comeback with intent:
to stand shoulder-to-shoulder with GPT-5 and Gemini 3 Pro — and perhaps, in some areas, to exceed them.

But the biggest test is still ahead:
How these models perform at global scale once rolled out to more users.

For now, one thing is clear — the AI race just got far more interesting.