New Large Generative AI Models Face Continuous Delays. Have They Reached Their Peak?

GPT-5 was expected to be available earlier this year. However, OpenAI released GPT-4.5 instead. This generative AI model was supposed to represent a significant advancement over its predecessors, but it ended up disappointing users. As a result, the company announced it would remove GPT-4.5 from its API in July due to its high costs and lack of return on investment.

This was already a troubling sign for the future of AI, but there’s more to the story.

What about GPT-5? GPT-5 was expected to arrive mid-year. However, after OpenAI CEO Sam Altman had been generating hype for months, its development proved to be problematic. The expected leap in performance hadn’t been realized, and the costs associated with developing it were immense.

As a result, OpenAI decided to delay its release and roll out GPT-4.5 instead. Still, GPT-4.5 has become one of the biggest disappointments in the company’s history. This raises significant concerns for the future of AI development.

Behemoth delayed. According to The Wall Street Journal, Meta will postpone the launch and deployment of its most ambitious model to date, Llama 4 Behemoth. This model boasts 288 billion active parameters (with a total of two trillion) and is the third iteration in the newly introduced Llama 4 family. However, the outlet reports “company engineers are struggling to significantly improve” its capabilities. The model was originally set to launch in April but is now estimated to arrive in the fall or even later.

Frustration. Sources close to the company told the outlet that management is frustrated with the performance of the team developing Llama 4 Behemoth. It’s already considering “significant management changes,” which could result in internal restructuring and possibly layoffs due to unsatisfactory results. Moreover, users aren’t receiving the currently available Llama 4 models as expected, marking another concerning sign.

Disbanding of the team. The Wall Street Journal also highlights that the first version of Llama was created by Meta’s Fundamental AI Research team, which consisted of academics and researchers. Since then, 11 out of 14 researchers have left the company.

More from Xataka On

AI Models Have Always Been a Big Black Box, Preventing Users From Knowing How They ‘Think’ Inside—Until Now

Anthropic isn’t making progress either. Claude, Anthropic’s generative AI chatbot, was expected to make a significant advancement. In February, the company unveiled Claude 3.7, which introduced some impressive features. However, Anthropic hasn’t launched Claude 3 Opus, considered the most ambitious model. Additionally, there’s no update on Claude 4.0. Again, this isn’t a good sign.

Small progress. There haven’t been major breakthroughs in model capabilities in recent months, but rather notable improvements in specific areas or flashy features. For instance, Gemini 2.5 Pro has shown remarkable strength in programming, which has allowed Google to make strides. Similarly, OpenAI gained attention with Studio Ghibli-style images, and Grok 3 became better known for its lack of censorship rather than its accuracy.

Slowdown. There’s renewed debate about a potential slowdown in AI development. The current approach of scaling (using more GPUs and larger datasets to train models) doesn’t seem to be delivering expected results as effectively. Delays on new large models raise questions about the future progress of generative AI.

Agents and reasoning AI. Models with reasoning capabilities have led to significant advancements in certain areas. Companies are beginning to introduce them along with in-depth research modes for specialized applications.

Another major hope for 2025 is the emergence of AI agents capable of autonomously completing sequences of tasks to solve problems, even integrating with other services or data sources. While there are impressive examples in programming, practical applications for end users are still limited at this stage.

Image | OpenAI

Related | Generative AI Creates a Gap Between Local and Cloud-Based Models. There’s Room for Both