OpenAI’s Latest Model Faces Scrutiny

In the fast-evolving world of artificial intelligence, OpenAI’s release of GPT-5 has sparked intense debate among developers and tech experts. Billed as the company’s most advanced model yet, GPT-5 promised breakthroughs in coding and agentic tasks, but early tests reveal a more nuanced picture. According to a hands-on review by David Gewirtz in ZDNet, the model’s coding capabilities fell short of expectations, prompting the tester to revert to GPT-4o for practical work. Gewirtz detailed attempts to generate code for tasks like creating a WordPress plugin and handling data visualization, where GPT-5 produced incomplete or erroneous outputs, requiring multiple iterations to fix.

This sentiment echoes broader feedback from the developer community. On platforms like X, users have expressed frustration, with one post noting that building with the GPT-5 API feels “frustrating af” due to difficulties in identifying failures compared to models like Claude’s Sonnet 3.5 or even GPT-4o. OpenAI’s own announcement on their blog highlighted GPT-5’s strengths in instruction-following and benchmarks like COLLIE and Scale MultiChallenge, yet real-world coding scenarios seem to expose limitations.

Benchmark Wins Versus Practical Shortfalls

Benchmarks paint GPT-5 as a leader, with reports from Vellum.ai showing strong performance in health-related tasks and general metrics. However, when it comes to coding, the model struggles with consistency. In Gewirtz’s ZDNet tests, GPT-5 failed to deliver a functional plugin on the first try, hallucinating features and ignoring specifications, whereas GPT-4o handled similar prompts more reliably. This gap suggests that while GPT-5 excels in controlled evaluations, its application in dynamic coding environments lags.

Comparisons with GPT-4o are particularly telling. A detailed benchmark from Passionfruit pitted GPT-5 against its predecessor, finding marginal improvements in reasoning but underwhelming results in coding efficiency and error rates. Users on X have corroborated this, with posts describing GPT-5’s outputs as “slop with random bolding,” lacking the polish of GPT-4o. Fortune’s coverage in their article notes reduced hallucinations and new “vibe coding” features, yet these innovations don’t fully compensate for practical deficiencies.

Developer Experiences and Industry Implications

Industry insiders are weighing in on these discrepancies. Simon Willison’s blog describes using GPT-5 as a daily driver, praising its steerability but acknowledging coding hiccups in complex scenarios. Similarly, a Hacker News thread linked from Y Combinator discusses the competitive clustering of AI models, with some researchers skeptical of GPT-5’s purported leaps. Posts on X highlight ongoing challenges, such as GPT-5’s “one shot laziness,” where it underperforms without extensive prompting, contrasting with GPT-4o’s robustness.

The pricing strategy adds another layer. TechCrunch reported in their piece that GPT-5’s low costs could ignite a price war, making it attractive despite flaws. Yet, for coders, value lies in reliability. Mashable’s take on vibe coding calls it a “dream come true,” but ZDNet’s tests suggest it’s more aspirational than operational.

Looking Ahead: Refinements and Expectations

OpenAI’s integration with Microsoft, as detailed in Microsoft’s announcement, promises broader access, potentially accelerating fixes through user feedback. Nathan Lambert’s analysis on Interconnects argues that while GPT-5 underdelivers on hype, its foundational advancements are phenomenal. X users, however, remain cautious, with some sticking to GPT-4o amid reports of degradation in earlier models.

Ultimately, GPT-5’s coding prowess may improve with updates, but current reviews indicate it’s not yet the game-changer promised. For industry professionals, this underscores the need for rigorous testing beyond benchmarks, ensuring AI tools align with real-world demands. As competition intensifies, OpenAI faces pressure to refine these capabilities swiftly.