On August 7, OpenAI launched its new flagship AI model, GPT-5, which is positioned as the first “unified” neural network combining both reasoning mode and quick response features, reports Baltimore Chronicle, with reference to ForkLog.
The CEO of the startup, Sam Altman, described GPT-5 as “the best model in the world” and “a significant step” toward creating artificial general intelligence (AGI), which could surpass humans in the most economically valuable tasks.
The model’s launch was met with criticism due to its incorrect answers, which led to user disappointment. GPT-5 produced slow, robotic responses that were short and unsatisfactory. Yuchen Jin, co-founder of Hyperbolic Labs, called the neural network a failure, noting that it still suffers from hallucinations, excessive use of dashes, and an inability to follow instructions properly.
In one instance, when GPT-5 was asked how many “b”s there were in the word “blueberry,” the model initially answered “three,” but later corrected itself after a follow-up query. Additionally, some developers noted that GPT-5 seemed to regress in basic programming skills, raising concerns about its reliability as an intellectual agent for autonomous coding.
Another point of dissatisfaction was the lack of choice for previous models. OpenAI removed support for earlier versions of ChatGPT, and later reinstated GPT-4o for Plus subscription users after backlash. Altman explained the poor response quality at launch was due to a broken router, which had been responsible for determining whether to apply the “thinking” or regular mode.
The model underwent several updates, including increased limits for users and the ability to select between two GPT-5 modes. Additionally, Altman promised to enhance the transparency of the answer search process, allowing users to see whether the “thinking” mode is being applied.
After these fixes, the model became more accurate and concise in its answers. It transitioned to the point more quickly, which professionals appreciated, though some users noted that the responses became more “dry” and less personable. According to one user, GPT-5 now better serves those who want a fast, clear answer, rather than those who enjoy long, friendly conversations with AI.
Regarding accuracy, OpenAI claimed that GPT-5 is much less prone to hallucinations. According to user feedback and tests conducted by ForkLog, this is true. GPT-5 has become more reliable in answering factual questions, and it more frequently responds with “I don’t know” when unsure. Users also noted that GPT-5 adheres better to instructions and maintains consistency in tone and style.
Despite these improvements, the model sometimes makes logical errors, reaching incorrect conclusions or offering inaccurate solutions. While the number of mistakes has decreased, issues with context and superficial responses persist.
GPT-5 also received improvements in programming, handling code faster and more accurately than previous versions. It demonstrates a deeper understanding of code and maintains coherence in long code fragments, a task that GPT-4 struggled with.
In terms of multimodal capabilities, GPT-5 has become more integrated, allowing it to process text, images, and voice within a single model. This major upgrade enables the model to generate more natural speech and analyze visual data with greater precision.
Despite the numerous improvements, some users still feel that GPT-5 is just a step forward, not a true revolution. According to them, the model didn’t meet the high expectations set by OpenAI, which may have oversold it.
Earlier we wrote that ChatGPT conversations may be revealed in court, warns OpenAI CEO.