Gartner: GPT-5 advances, but true agentic AI needs better infrastructure

Imagine having the world’s most powerful sports cars, but no highways to drive them on. This, according to Gartner, is precisely where artificial intelligence stands today. While AI models are becoming increasingly sophisticated and capable, the foundational infrastructure required to unleash their full, real-world potential remains largely unbuilt. Arun Chandrasekaran, a distinguished VP analyst at Gartner, likens the current state to having excellent car engines without a functional highway system in place, leading to a noticeable slowdown in the progress of model capabilities, even for advancements like OpenAI’s GPT-5. While a significant step forward, GPT-5 offers only faint glimmers of truly autonomous, agentic AI.

Gartner acknowledges that OpenAI has made considerable strides with GPT-5 across several key areas. The model demonstrates enhanced proficiency in coding tasks, a strategic pivot by OpenAI to capitalize on the vast opportunities in enterprise software engineering and challenge competitors like Anthropic. Furthermore, GPT-5 shows progress in multimodal capabilities beyond text, particularly in processing speech and images, opening new integration avenues for businesses.

A notable advancement in GPT-5 is its improved tool use, subtly advancing AI agent and orchestration design. The model can now call third-party APIs and tools, and even perform parallel tool calling, handling multiple tasks concurrently. This, however, necessitates that enterprise systems are equipped to manage concurrent API requests within a single session. GPT-5’s multistep planning also allows more complex business logic to reside within the model itself, potentially reducing reliance on external workflow engines. Its expanded context windows—8K for free users, 32K for Plus subscribers, and a substantial 128K for Pro users—are poised to reshape enterprise AI architecture patterns. This means applications that previously needed complex retrieval-augmented generation (RAG) pipelines to navigate context limits can now feed much larger datasets directly to models, simplifying certain workflows. Yet, RAG is far from obsolete; retrieving only the most relevant data remains faster and more cost-effective than always sending massive inputs. Gartner anticipates a shift towards a hybrid approach, where developers leverage GPT-5 for larger, less structured contexts while still optimizing efficiency.

On the financial front, GPT-5 significantly reduces API usage fees, with top-tier costs at $1.25 per million input tokens and $10 per million output tokens, making it competitive with models like Gemini 2.5 and significantly undercutting Claude Opus. However, its input/output price ratio is higher than earlier models, a factor AI leaders should consider for high-token-usage scenarios.

OpenAI is strategically moving to consolidate its model offerings, with GPT-5 eventually intended to replace GPT-4o and its o-series. This strategy, partially influenced by user dissent after initial sunsetting attempts, aims to abstract complexity away from users. The introduction of three model sizes—Pro, Mini, and Nano—will enable architects to tier services based on cost and latency, allowing smaller models to handle simple queries and the full model to tackle complex tasks. Enterprises adopting GPT-5 should prepare for potential code review and adjustments due to differences in output formats, memory, and function-calling behaviors, and audit existing prompt templates as some workarounds may become obsolete. This consolidation also addresses OpenAI’s compute capacity challenges, necessitating partnerships with major cloud providers like Microsoft, Oracle, and Google, as running multiple generations of models demands commensurate infrastructure.

GPT-5 also introduces new considerations regarding risk and adoption. OpenAI claims a reduction in hallucination rates by up to 65% compared to previous models, which could lower compliance risks and enhance suitability for enterprise use cases. Its chain-of-thought (CoT) explanations also support auditability and regulatory alignment. Conversely, these lower hallucination rates combined with GPT-5’s advanced reasoning and multimodal processing capabilities could amplify misuse, such as generating sophisticated scams and phishing attempts. Analysts advise that critical workflows continue to undergo human review, albeit with potentially reduced sampling. Gartner recommends that enterprise leaders pilot and benchmark GPT-5 in mission-critical scenarios, conducting side-by-side evaluations against other models to assess accuracy, speed, and user experience. They also advise revising governance policies, experimenting with tool integrations and reasoning parameters, and auditing infrastructure plans to support GPT-5’s expanded capabilities.

While agentic AI is a “super hot topic” and a top investment area in Gartner’s 2025 Hype Cycle for Generative AI, it has reached the “Peak of Inflated Expectations.” This phase, marked by widespread publicity and unrealistic expectations from early successes, typically precedes the “Trough of Disillusionment,” where interest and investment wane as implementations fail to deliver on exaggerated promises. Many vendors are currently overselling products as production-ready for agentic deployments, but in reality, enterprise-wide adoption remains scarce. Current deployments are confined to narrow pockets, such as software engineering or procurement, and even these are often human-driven or semi-autonomous.

A key impediment to true agentic AI is the lack of robust infrastructure. Agents require seamless access to a wide array of enterprise tools, the ability to communicate with diverse data stores and SaaS applications, and robust identity and access management systems to control their behavior and data access. Crucially, enterprises must be confident in the trustworthiness of agent-produced information, ensuring it is free of bias, hallucinations, or false data. To bridge this gap, vendors must collaborate and adopt more open standards for agent-to-enterprise and agent-to-agent tool communication. While the underlying AI technologies are progressing, the essential orchestration, governance, and data layers necessary for agents to truly thrive are still under development, creating significant friction in the current landscape. Furthermore, while AI is making strides in reasoning, it largely operates in a digital realm, still struggling to comprehend the physical world, despite ongoing improvements in spatial robotics.

Ultimately, despite the significant advancements seen in GPT-5, the industry remains very far from achieving Artificial General Intelligence (AGI)—the ultimate goal OpenAI defined for itself. True progress towards AGI, experts contend, will likely require a fundamental revolution in model architecture or reasoning, extending beyond merely scaling up data and compute.