The LLM Landscape Shifts as Google Rises, Meta Stumbles, and OpenAI Rethinks Strategy
In the early days of large language models (LLMs), OpenAI led the pack, closely followed by Meta. OpenAI’s GPT models set the standard for performance, while Meta established a strong presence with its open-weight alternatives—models with freely accessible code that anyone can deploy or modify. Meanwhile, companies like Google, despite pioneering the transformer architecture in 2017, were left trailing. Instead they are recalled more for the clumsy rollout of Bard in 2023 than their foundational AI work.
But the dynamics are changing. New developments from Google and setbacks at both Meta and OpenAI are reshaping the competitive landscape.
Also read: Zoom Unveils Major AI Upgrade Across Workplace Suite
Meta’s Llama 4 Faces a Rough Start
Meta’s release of Llama 4 came unexpectedly on a Saturday, 5 April. The timing surprised many, effectively burying the announcement under the next week’s news cycle. While Llama 4 introduced impressive capabilities, including multimodal support and multiple model sizes—Behemoth, Maverick, and Scout—critics quickly raised concerns.
The Scout variant drew attention for its massive 10-million-token context window, far exceeding the norm of around one million tokens. However, benchmark testing showed that despite the technical specification, Llama 4 lagged behind rival models in long-context performance.
Controversy deepened when it was revealed that Meta had used a specially tuned version of Llama 4 to rank on the LMArena leaderboard, rather than the general release version. This tactic raised questions about transparency. Furthermore, Meta held back the release of smaller variants and a dedicated “reasoning” model, though it has promised the latter is forthcoming. Observers noted that the release appeared rushed and incomplete.
OpenAI Pulls Back GPT-4.5
OpenAI also encountered headwinds. GPT-4.5, introduced in February 2025 as a research preview, was billed as its most powerful chat model yet. While it outperformed previous iterations in many benchmarks, its API pricing—$150 per million output tokens—drew sharp criticism. This rate was 15 times higher than GPT-4o, priced at just $10 per million tokens.
Estimates suggest GPT-4.5 used a mixture-of-experts architecture with roughly 5.4 trillion parameters, making it one of the largest LLMs released to date. However, such scale came with steep hardware and operational costs.
OpenAI announced it would retire GPT-4.5 from API access, less than three months after launch. The model remains available through the ChatGPT interface. To replace it, OpenAI introduced GPT-4.1, a more cost-effective option at $8 per million tokens. Though it doesn't match GPT-4.5 in every metric, it performs better in some programming-related tasks.
OpenAI also launched new reasoning models—o3 and o4-mini. The o3 model, in particular, performed well in benchmark tests, but once again, pricing raised eyebrows: $40 per million tokens via API.
Google’s Gemini Surges Ahead
The setbacks at Meta and OpenAI created an opportunity—and Google seized it. Models such as DeepSeek-V3, Alibaba’s Qwen2.5, and Google’s own Gemma quickly gained traction. These open-weight models now dominate platforms like Hugging Face and LMArena, offering performance that rivals or exceeds Llama 4 at a much lower cost.
But, it is Google’s Gemini 2.5 Pro that takes away the spotlight. Released on 25 March, this multimodal, “thinking” model can reason through complex tasks using self-prompting. With a one-million-token context window and deep research capabilities, it has quickly topped leaderboards like SimpleBench, the AI Intelligence Index, and LMArena.
Gemini 2.5 Pro is also notably affordable. Google offers free access through its Gemini app and AI Studio. API pricing is also competitive. It cost $10 per million tokens for Gemini 2.5 Pro and just $0.40 per million tokens for Gemini 2.0 Flash.
For those handling large-scale workloads, developers are increasingly turning to Google or DeepSeek models over OpenAI, citing cost-effectiveness and strong reasoning capabilities.
A Changing Tide in the LLM Race
Despite the challenges, neither OpenAI nor Meta is out of the picture. OpenAI’s ChatGPT remains a massive force with a reported user base of one billion. Meta continues to push updates to Llama 4 and promises new capabilities soon.
But the momentum appears to have shifted. Google’s Gemini line, backed by strong performance, competitive pricing, and free availability, is emerging as the new standard in the LLM space. As innovation continues, the balance of power in AI may soon look very different from what it was just a year ago.
