If you follow the evolution of LLMs even slightly, you’ve likely noticed the constant shifts and developed your own personal usage preferences. Looking at the chart below (from LMArena – Goldman Sachs Global Investment Research), the changes are undeniable. The Goldman Sachs data is brutal: what was once a cutthroat “knife fight” between startups has turned into a Google monologue.
2024: The Year of Sleepless Nights and Musical Chairs
To understand the shock of 2025, we must recall how frantic 2024 truly was. It was the year of the “War of Attrition”:
- Anthropic shocked the world in March with Claude 3 Opus, becoming the first to dethrone OpenAI’s nearly year-long reign.
- OpenAI didn’t take it sitting down, reclaiming the top spot with GPT-4o in May, only to be challenged again by Claude 3.5 Sonnet in June.
- We ended the year with the rise of “reasoning models” (o1), as OpenAI desperately tried to hold onto its crown.
In 2024, the top of the LMArena was like a high-turnover hotel. No leader could sleep soundly.
2025: Google’s “Shop Floor” Swept the Chart
Now, observe the blue wave dominating almost 100% of the right side of the image. What happened?
Google, which spent much of 2024 being labeled as “lagging,” decided to get serious with the return of co-founder Sergey Brin to the front lines. The result was an unprecedented territorial occupation. While OpenAI and Anthropic accumulated glory days in the past (represented by OpenAI’s 540 days in green), in 2025 they were pushed to the bottom of the podium.
Curious Facts from 2025:
- Dominant Gemini: Google held the top spot for over 90% of the days in 2025. That’s 302 days (and counting) of absolute sovereignty.
- The Grok “Intruder”: The only company that managed to break through Google’s blockade—if only for 34 days—wasn’t OpenAI, but Elon Musk’s xAI. Grok was the only breath of variety in a year painted blue.
- OpenAI and Anthropic in “Almost” Territory: They remain excellent, but the chart doesn’t lie: in terms of being “the best model in the world” on LMArena, they have lost their leadership momentum.
The question remains: Has Google finally found the formula for permanent hegemony, or is OpenAI just stockpiling ammunition for a historic counter-attack? If 2024 was the year of the struggle, 2025 is the year of dominance.
Important Add-on: How did xAI dethrone Google for 34 days?
Looking at the chart, the small “gray wave” in the middle of Gemini’s blue ocean might look like a statistical error, but it was one of the most talked-about moments of 2025. How did a company so much younger than Google reach the top of LMArena?
While Google optimized Gemini for efficiency at scale, Elon Musk activated Colossus, the world’s largest GPU cluster.
- Grok-3 (the model likely responsible for this peak) was trained with unprecedented computational power in a very short timeframe.
- On LMArena, this translated into a model that rarely “hallucinated” during complex logic tasks, beating Gemini in pure coding and heavy mathematical reasoning during its launch month.
Furthermore, LMArena rankings are based on human preference (Crowdsourcing).
- Google is known for being extremely cautious (and sometimes “stifled”) regarding safety and bias.
- Grok, on the other hand, was tuned to be more direct and witty. In blind tests, many users preferred Grok’s answers simply because they felt less “robotic” and more assertive than Gemini’s, quickly boosting its Elo Score.
The Bottom Line: Grok proved that with massive compute and a less conservative approach, it is possible to unseat giants. But the chart is clear: getting to the top is one thing; staying there is Google’s game.
And how are the Chinese models doing on LMArena?
While the 2025 chart shows a visual battle between Gemini, OpenAI, and Grok, there is a silent force gaining ground: Chinese models. Names like Qwen and DeepSeek are no longer just “promises.” In 2025, they dominated LMArena’s Programming and Mathematics categories, proving that U.S. dominance is now challenged not by a single company, but by an entire ecosystem from across the globe.
| Feature | Google Gemini (2025 Leader) | Qwen / DeepSeek (China) |
| LMArena Strength | Multimodality (Video/Audio) and Giant Context Window. | Logical Reasoning, Math, and Coding. |
| Response Style | Polished, safe, and highly informative. | Direct, technical, and fewer Western safety “filters.” |
| Availability | Closed ecosystem (Google Cloud/Vertex). | Open-weights models for the community. |
| Value for Money | High (focused on enterprise/premium users). | GPT-4 level performance at a fraction of the price. |
If the Goldman Sachs chart shows that Google won the battle for time at the top, LMArena also tells us that China won the battle for the democratization of power.