OpenAI’s much-anticipated GPT-5 AI model was released on Aug. 7 with a resounding thud. After it went live, a Reddit thread titled ”GPT5 is horrible” quickly amassed more than 6,000 upvotes, while OpenAI CEO Sam Altman and other senior employees faced numerous complaints about GPT-5 during an ”ask me anything” (AMA) session. Users called the new model a downgrade and demanded access to older AI models that were removed upon GPT-5’s release. Some also complained that GPT-5’s responses were too direct, lacking the conversational personality found in previous versions. All of this backlash was despite GPT-5 and GPT-5 Pro showing improvements over earlier models on most AI-grading benchmarks. Nevertheless, within 24 hours, Altman acknowledged users’ complaints and announced that most of the legacy models initially disabled with GPT-5’s release were available again.

It is common for new generations of technology to offer diminishing returns beyond the first few releases. This is because innovation often follows an S-curve, where initial development and adoption is slow, followed by rapid growth, and then a plateau as the technology matures. Early smartphones, for example, delivered remarkable advancements in processing power and internet connectivity speeds, making each annual release of the new iPhone model feel like a significant leap. However, the breakneck speed of technology innovation today means this cycle is happening faster than ever. And for AI chatbots focused on generating basic text responses, the disappointment surrounding GPT-5’s release suggests that new iterations, even if improved, are already entering this final plateau phase.

Doomed From the Start?

In hindsight, the widespread disappointment with GPT-5 was fairly predictable. For simple, one-shot prompts (such as those requesting a quick draft email or fact recall), advancements in AI chatbots like ChatGPT have shown diminishing returns for at least a year. The progression from GPT-2, which often produced basic and nonsensical conversational replies, to GPT-3, which was prone to rapid and frequent hallucinations, and finally to GPT-4 and GPT-4o, which offered far more coherent answers, represented substantial improvements. But since OpenAI’s previous frontier model was already so effective, GPT-5 was never going to be able to show as generational a leap, at least for the simple prompts and conversations that the vast majority of its users leverage it for. For those users who came to love GPT-4o’s style and conversational fluency, GPT-5’s deviation in tone and the lack of a significant conversational leap meant it was always doomed to be a letdown.

However, the next leap won’t come from more refined conversations with AI, but from AI systems gaining agency — including the ability to plan, coordinate and execute tasks through multiple AI agents. GPT-5’s advancements over previous OpenAI models, particularly in coding and web design, are an example of this. Its ”thinking mode” excels at complex, multi-step analysis, enabling more sophisticated responses beyond simple text or image generation. GPT-5 is also far better at writing Python scripts, creating webpages and applications, and processing documents, datasets and other files to carry out analysis. Competitors like Google’s frontier model Gemini 2.5 Pro and Anthropic’s frontier model Claude Opus 4.1 are becoming more adept at integrating various specialized AI capabilities to build more complex outputs as well.

But making this shift will require addressing the growing training costs of AI models, which are now reaching $1 billion for frontier models, as well as growing costs of inference (i.e., querying a trained model), both in terms of dollars and electricity. Indeed, just days before GPT-5’s release, OpenAI unveiled two open-weight language models that users can install on their own devices: gpt-oss-120b and gpt-oss-20b. The company claims that gpt-oss-120b, a 117 billion parameter model, is almost as powerful as its older o4-mini model and can operate on a single 80 GB GPU (i.e., one Nvidia A100 AI chip, which costs about $10,000). OpenAI also claims that gpt-oss-20b, a 21 billion parameter model, is nearly on par with its o3-mini model and can even run on some gaming laptops, requiring only a 16 GB GPU. Businesses will be attracted to more affordable and efficient models, whether run locally or via an API from a company like OpenAI. This is because many simple use cases involve generic queries, meaning the advantages of a more costly model aren’t always beneficial when integrating with a company’s own products and specialized systems that handle their proprietary data.

The Rise of Agentic AI

Effectively, AI innovation is evolving beyond training simple chatbots toward building more robust systems that can layer on different capabilities, known as agentic AI. These systems will be able to leverage AI, often based on the large language models that power chatbots, to unify them or coordinate different AI agents, such as those that process image inputs (like viewing a webpage) and those that control mouse clicks on a computer. Agentic AI has emerged as the new focus for AI enthusiasts and experts over the past 18 months — largely superseding chatbots’ strictly text responses, which comparatively offer less obvious and valuable use cases. While AI chatbots can answer questions and generate images, a fully agentic AI system, or one that layers on various AI agents, can actually achieve an end goal or accomplish a specific task by autonomously determining the most effective approach through the coordinated use of different AI agents.

Thus, while improvements in these areas in models like GPT-5, Gemini 2.5 Pro and Claude Opus 4.1 may not be apparent to most users, they are far more significant for power users — i.e., companies and individuals focused on more advanced applications. The ”wow” factor now stems from the expanded scope and speed of new AI systems, such as their ability to rapidly analyze codebases, accurately draft and debug complex code, and process hours of audio.

Another key area where AI developers are looking to expand capabilities is around the so-called context window or context length, which refers to the amount of information an AI system can process at once, measured in tokens. Most of today’s AI models can handle 32,000, 64,000, or 128,000 tokens. To put this in perspective, one token is about four characters or three-quarters of a normal word. Thus, 32,000 tokens translate to about 24,000 words of text, roughly 50 pages of single-spaced text, and a few thousand lines of code. Expanding this capacity is crucial for tackling more complicated tasks, such as processing thousands of pages of corporate information, writing extremely concentrated code, and completing advanced agentic AI tasks where the model’s oversight is limited by its context window.

One of the biggest complaints by more technologically proficient users of GPT-5 has been that the free and cheaper tiers of the new model only have 16,000 and 32,000 token context windows, respectively. This is a significant downgrade from ChatGPT’s previous o3 and o4-mini models, which both had a 200,000 token context length window, making them far more useful for processing larger documents. Amid this criticism, OpenAI competitor Anthropic announced that its Claude Sonnet 4 model, when accessed via the company’s API, had a 1 million window token length. However, Anthropic’s most advanced chatbot model, Claude Opus 4.1, maintains a 200,000 token context length.

The most innovative and complex AI systems place even greater demands on hardware, especially for inference. This is because they layer multiple AI prompts or queries on top of each other to generate sophisticated responses and perform increasingly useful tasks. Due to this escalating need for computation and hardware, the biggest advancements are increasingly being done behind paywalls or are accessible only via APIs instead of chatbots, meaning they are less visible to most users who do not need them. This trend also necessitates increased investment, development and deployment of hardware to run these systems, such as more capable AI accelerators and larger, more power-hungry data centers — even if the utility of chatbots for the majority of users is reaching a plateau. But while these more task-oriented AI applications are more costly to develop and operate, they may be precisely what AI developers need to generate substantial revenue, moving beyond the single-seat chatbot users paying $20 or $25 per month.

The Bigger Picture

Although chatbot advancements may be plateauing, the broader field of AI development is not slowing down anytime soon. More complex applications, including those relevant to business and national security, remain in the rapid innovation phase of the aforementioned S-curve. From a geopolitical perspective, the focus on AI as a strategic technology and a strategic risk for societies will thus remain crucial.

The almost exponential-like scaling nature of using more complex agentic AI systems intensifies dependency risks. Chinese AI developers, for example, are reliant on H20 chips produced by the U.S. company Nvidia for the inference capacity of their models, rather than Ascend chips made by the Chinese company Huawei. If the United States restricts China’s access to those Nvidia chips in the future, Chinese developers and China would face national security and capacity concerns about scaling up these more business- and geopolitically-relevant AI systems, because they would need to rewrite entire codebases to flip from using Nvidia’s chips to Huawei’s chips. To mitigate this risk, China will increasingly push for Chinese companies to use Huawei and other domestic chips over foreign alternatives like Nvidia.

The European Union’s Artificial Intelligence Act, still in its implementation phase, may also not fully address the complexities of fast-scaling and horizontally integrated AI agents overseen by single or multiple AI models, which collectively form a complete agentic AI system. This oversight stems from the fact that the AI regulations were drafted largely before and in the early stages following ChatGPT’s release. For example, the act’s high-risk designation for AI models primarily focuses on regulating individual AI models, not systems of AI models, automatically labeling those with a certain total compute in training as high risk. Consequently, as more resource-intensive AI applications emerge, requiring increased inference and data center capacity, the United States is poised to maintain or even expand its lead in AI development. The European Union and China, in contrast, face constraints due to limited AI industry and hardware manufacturing capabilities, respectively.

Moreover, these more advanced models are unlikely to dispel fears that AI will cause a significant overhaul of the workforce and jobs that humans perform. The advancements in AI’s ability to code, as well as advancements in using AI agents or agentic AI to automatically structure unstructured data, perform data entry or summarize vast amounts of information at one time, are all things that will have a major impact on the workforce and are tasks that many workers perform on a daily basis.

Additionally, more advanced AI models are unlikely to ease concerns about the technology’s potential to eradicate jobs that humans currently perform. The advancements in these systems’ ability to code, perform data entry, use AI agents to structure unstructured data and summarize vast amounts of information simultaneously will significantly affect the job market, as these are tasks many workers perform on a daily basis. Indeed, evidence already points to a ”hollowing out” of entry-level jobs and internships that involve these tasks — roles that traditionally provided on-the-job training for entry-level workers to develop into more skilled professionals. Currently, very few governments have taken significant steps to address or even comprehend the overhaul of the job market and workforce brought about by the advanced systems that AI developers have shifted toward. This inaction keeps the risk of societal disruptions and upheaval high, even if AI doesn’t ultimately take everyone’s jobs, as the most pessimistic critics warn.

So while GPT-5’s debut may leave casual users wanting more, it signals the maturity of today’s AI chatbots — and the rise of new AI systems that are set to drive the global AI arms race and disrupt job markets.