Redazione RHC : 9 August 2025 08:55
OpenAI has launched its new flagship model, and the message is clear: reliability, power, and a radical change in the way we interact with artificial intelligence. After years of viewing language models as experimental tools, GPT-5 marks the definitive transition from prototype to production.
One of its strengths is the unified system with “smart router”: there’s no longer any need to manually choose which model to use. GPT-5 analyzes the request and autonomously decides whether to provide a rapid response or activate a more complex computation mode for challenging problems. A single interface, power on demand.
On the reliability front, progress is remarkable. OpenAI claims a reduction in hallucinations of between 45% and 80% compared to previous models, the result of targeted engineering work. This is a crucial step towards making AI a practical, safe, and predictable work tool.
Performance on key tasks confirms the leap in quality: in coding, GPT-5 achieves a 74.9% score on SWE-bench, placing it among the top performers in software engineering; in scientific reasoning, it shows significant progress in mathematics (AIME), science (GPQA), and multimodal perception; In writing, it demonstrates superior management of structure, rhythm, and metaphor, going far beyond simple grammatical correctness.
Finally, access and costs: GPT-5 is already available in ChatGPT (replacing GPT-4o) and via API in various formats (-5, mini, nano) with 256k input and 128k output context. The rates, up to $10 per million output tokens and $1.25 per million input, are extremely competitive, in line with models like Google Gemini 2.5. But the real breakthrough isn’t just power: it’s trust. The age of toys is over—this AI is designed to work.
OpenAI presented GPT-5 as an infrastructure model: reliable, powerful, and designed for professional use. Now let’s add a surprising aspect: its development would have required a billion-dollar investment and unprecedented hardware infrastructure, confirming that this is no longer a prototype, but a full-fledged undertaking.
GPT-4 is estimated to have cost around $63 million to train, using 25,000 Nvidia A100 GPUs for 90-100 days. By comparison, estimates for GPT-5 suggest a cost of $1.25 to $2.5 billion and an infrastructure of 250,000 to 500,000 dedicated Nvidia H100 GPUs. Alternatively, according to other calculations based on market estimates (parameters of 8.8-15 trillion and training durations around 150 days), between 150,000 and 260,000 H100 GPUs would be needed, for a hardware cost of between $3.7 and $6.6 billion just to purchase the cards.
While official details on GPT-4’s parameter count have not been made public, sources estimate it could reach 1.8 trillion parameters. GPT-5, on the other hand, is projected to have between 8.8 and 15 trillion parameters, which is about 5-8 times larger than its predecessor.
The required infrastructure is not only expensive, but It’s also logistically challenging: hundreds of thousands of H100 GPUs represent a massive demand for both energy and physical space. If GPT-4 had already marked a turning point in resource use, GPT-5 pushes it to near-industrial levels, effectively embodying the definitive shift from experimentation to AI infrastructure.
From a practical standpoint, GPT-5 promises far greater power and reliability than GPT-4. However, the high cost and enormous hardware requirements shift the focus: we are no longer in the realm of experimentation, but in a new era in which AI is infrastructure—expensive, widespread, and ready to serve mission-critical applications. The age of toys is over: AI is now designed for work.