Photograph: Open AI/Youtube
At the peak of hype, even a small step forward can feel like a stumble. OpenAI’s release of GPT-5 was heralded as a milestone on the march to artificial general intelligence (AGI), a point at which machines might outstrip humans across cognitive tasks. Rather the model’s underwhelming debut has provoked a more pragmatic question: what if the future of artificial intelligence looks a lot like the present?
The new system was pitched as a smoother, more refined version of ChatGPT, the conversational agent that made OpenAI a household name. Internally executives spoke of good “vibes”, and expected improvements in usability and fluency. Externally the reception was muted. Images circulated of the model making familiar factual errors. Developers expressed nostalgia for earlier versions. Benchmarks placed GPT-5 comfortably within the pack rather than ahead of it.
The launch has unsettled assumptions that had become doctrine in Silicon Valley. Since 2020 a single formula has dominated the field: pour more data and computing power into transformer models and expect exponential returns. GPT-5, reportedly trained on hundreds of thousands of Nvidia’s cutting-edge processors, appears to have tested the boundaries of that formula. It did not break through.
For OpenAI the implications are commercial as well as technical. The company’s valuation, floated at $500bn, rests on the premise AGI is within reach and that OpenAI is leading the way. GPT-5’s performance has called both assumptions into question. Sam Altman, the firm’s chief executive, admitted although foundational models are improving, the user experience of ChatGPT may not see major leaps.
Competitors have taken note. Google, Anthropic and xAI have caught up in model performance. Meta meanwhile is betting on alternative approaches entirely. Its researchers argue models trained only on text are insufficient to build machines that understand the world. Instead they propose “world models” trained on video, spatial data and memory. These could prove more adaptable, particularly in areas such as robotics and simulation.
The slowdown is not merely conceptual. It reflects constraints in inputs. AI developers have exhausted publicly available high-quality training data. Firms are now signing licensing deals with publishers and other rights holders. At the same time, compute costs are rising steeply. GPT-5’s training alone may have consumed tens of millions of dollars in energy and hardware.
These pressures are shaping policy as well as product. In Washington the tone has changed. The Biden administration, concerned about existential risks, championed regulation and guardrails. The Trump administration is instead positioning AI as a strategic export. Recent decisions to permit Nvidia to resume chip sales to China, including variants of its high-end Blackwell line, suggest a move away from containment towards dominance.
David Sacks, the new AI adviser to the Trump administration, has declared predictions of AGI-induced job loss are as overblown as AGI itself. His framing of the market as balanced and manageable signals a wider reassessment of risk. American officials now speak less of intelligence explosions and more of supply chains and standards.
The business community has adjusted faster. Investors continue to pour capital into AI start-ups and infrastructure, undeterred by questions of diminishing returns. Nvidia’s valuation has risen by 25% in recent months. SoftBank’s share price has surged. OpenAI’s own revenues from ChatGPT have reached an annualised $12bn. The model’s linguistic adequacy, rather than its ingenuity, appears sufficient to support commercial success.
Start-ups are increasingly focused on application rather than breakthrough. OpenAI and others are embedding engineers within client firms, treating their models as components of business systems rather than autonomous minds. Academic researchers have begun systematically testing models on practical tasks such as coding, data retrieval and customer service. GPT-5 performs adequately and quickly even where it does not dominate benchmarks.
Some researchers believe the field now resembles the early internet: overcapitalised, underregulated, but structurally embedded. The potential for product innovation remains vast. Video generation, synthetic voice, autonomous agents and industry-specific tools are all areas of active development. The consumer market remains enthusiastic even as technologists become more circumspect.
None of this implies the pursuit of AGI has been abandoned. Masayoshi Son, head of SoftBank and one of OpenAI’s major backers, remains committed to the goal. Meta continues to invest heavily in research. But the strategic focus has shifted from acceleration to adaptation. The race is not to build the smartest machine, but the most useful one.
What GPT-5 has revealed is not a ceiling but a plateau. The rapid vertical gains of the past three years have given way to slower, more incremental advances. This is typical of maturing technologies. The steam engine, the aeroplane, the semiconductor all followed similar arcs. Periods of explosive growth yielded to phases of optimisation and integration.
The AI sector now faces that inflection. The myth of runaway intelligence has given way to the reality of managed complexity. GPT-5 may not inspire awe but it works. For companies deploying models across logistics, education or legal services, that is sufficient. For researchers chasing AGI it is a moment for reflection.
Silicon Valley continues to speak the language of disruption. But the most valuable AI companies are starting to behave like infrastructure providers. Their models are being embedded, licensed and scaled. The excitement around GPT-5 may have faded. The influence of the systems it represents is only deepening.
History suggests transformative technologies rarely unfold in linear fashion. They arrive in waves, shaped as much by economics as by science. GPT-5 may not mark the beginning of AGI. It may rather signal the start of the industry’s next phase: one in which utility, rather than vision, defines progress. ■