Opinion: Sam Altman’s GPT-5 promises crumble under dev tests

For the last few years, OpenAI promised that the GPT-5 model would be a major leap forward in AI development.

Ahead of last week’s launch, executives talked up improved “thinking” capabilities, faster coding, and more powerful autonomous agents. The company claimed GPT-5 was “state of the art.”

Image credit: Timmy Loen

Pfffft! That didn’t happen. The smoke rising from OpenAI’s disastrous launch hasn’t cleared yet.

But despite all the headlines and the drama around GPT-5, there is a new model out.

Once you get past the fuss around the model routing issue – OpenAI removed the option to manually select a model – you need to ask the important question: How smart is it?

After talking to developers and founders in Southeast Asia and the US, I knew that GPT-5 was not living up to the hype.

Some users found incremental improvements; others saw no difference at all. And after putting the model through my own tests, I can see why so many walked away underwhelmed.

Unimpressive first impressions

I spoke with developers and superusers to get a better sense of how GPT-5 was performing. Most of the people I spoke with – at least those willing to try it despite the bad press – were unimpressed, although no one said it was bad.

A lawyer tells me she had to write far more context into her prompts to get useful answers. “Otherwise, it was not much different from GPT-4,” she says.

A startup consultant in Singapore describes GPT-5 as “quicker and more focused,” but adds that it was neither “incredible nor groundbreaking.”

For the most complex tests – coding and reasoning – I turned to developers who live and breathe this stuff.

U-Zyn Chua, an adjunct lecturer at NUS-ISS and an independent AI engineer, tells me that the first few days with GPT-5 were “a mess: slow responses, inconsistent behavior, and logical questions that the GPT-5 Thinking model got wrong even after working on it for over two minutes.”

See also: Price beats intelligence for OpenAI’s newest model

Two plus two equals … something?

Agents: licensed to fail

Writer, rewrite thyself