GPT-5.4 Thinking Delivers Deeper Analysis, But Struggles with Specifics

The latest AI model from OpenAI offers more cognitive capabilities, but sometimes answers questions it wasn't asked.

Published on Mar. 9, 2026

Got story updates? Submit your updates here. ›

I put OpenAI's new GPT-5.4 Thinking model through a series of tests, including image generation, travel planning, and analysis of social media's impact. While the model delivered thoughtful and in-depth responses, it sometimes struggled to directly answer the specific questions I asked, instead providing tangential answers. The text-based responses were strong, but image generation and formatting lagged behind. Overall, GPT-5.4 Thinking shows promise, but requires careful management to keep it on track.

Why it matters

As AI models become more advanced, it's important to understand their capabilities and limitations. GPT-5.4 Thinking represents a step forward in cognitive abilities, but the tendency to answer questions that weren't asked raises concerns about relying on these models for critical tasks without close supervision. This story highlights the need for continued development and refinement of AI to ensure it can reliably follow instructions and provide relevant, on-point responses.

The details

In my tests, GPT-5.4 Thinking demonstrated strong reasoning and analysis skills, providing detailed responses on topics ranging from designing a futuristic aircraft carrier to the societal impact of social media. However, the model sometimes answered questions that differed from what I asked, requiring me to steer it back on track. Additionally, its image generation and formatting capabilities lagged behind the quality of the text-based responses. For example, when asked to generate an image of a flying aircraft carrier, the model simply produced the same flawed image it had generated previously, despite my detailed design specifications. The formatting of the responses also left much to be desired, with the model favoring long, numbered lists that were not always easy to parse.

  • On March 9, 2026, I tested the new GPT-5.4 Thinking model from OpenAI.
  • The GPT-5.4 Thinking model was released by OpenAI the previous week.

The players

OpenAI

An artificial intelligence research company that has developed a series of advanced language models, including the latest GPT-5.4 Thinking model.

David Gewirtz

A technology journalist who put the GPT-5.4 Thinking model through a series of tests to evaluate its capabilities and limitations.

Got photos? Submit your photos here. ›

What they’re saying

“I have often characterized ChatGPT as a bright college student in need of good supervision. I would characterize GPT-5.4 Thinking as a very bright grad student who definitely needs good supervision.”

— David Gewirtz (ZDNET)

“Whenever I see results like this, I get more and more concerned about a world overrun by AI agents. Yes, the AI may sometimes know better. Humans definitely need help. But I'd really like AIs to follow our instructions. I'm not ready to accept it as our AI overlord just yet.”

— David Gewirtz (ZDNET)

What’s next

OpenAI will likely continue to refine and improve the GPT-5.4 Thinking model, addressing the issues identified in this testing to ensure the AI can more reliably follow instructions and provide relevant, on-point responses.

The takeaway

While GPT-5.4 Thinking demonstrates impressive cognitive capabilities, the model's tendency to answer questions that were not asked raises concerns about relying on it for critical tasks without close supervision. As AI models become more advanced, developers must prioritize ensuring the models can reliably follow instructions and provide relevant, actionable responses to maintain trust and confidence in these technologies.