We Tested OpenAI’s New Models. Here’s Our Honest Take.

OpenAI released GPT-4.5 and GPT-4.5 Instant within the same week. We have both been putting them through their paces and wanted to give you a real account of what we found, where they fall short, and what they are actually worth using for.

The short version: mixed results, with one significant problem that makes the flagship model difficult to recommend for most everyday use cases right now.

The Speed Problem

The biggest complaint from both of us is the same: GPT-4.5 is slow. Not a little slow. Borderline unusable slow for anything that requires a fast turnaround.

James spent the weekend helping his daughter build out a website for a new business concept. He gave GPT-4.5 Pro a shot, thinking it would be a good test case. The experience was not what he hoped for.

“She would submit a prompt or a request and we had to leave and come back,” James said. “It would take 45 minutes or an hour and a half. It just does not seem to have the prompt intelligence that Claude does today.”

For context, the same kind of comprehensive request in Claude Code CLI, one that might involve changing features across an entire site, tends to take around 15 minutes. The equivalent in GPT-4.5 Pro came back in over an hour.

Sean had a similar experience across different use cases.

“The biggest problem I had was it takes such a long time that I almost feel like it is broken or hung up,” Sean said. “I’ll come back to it and think, oh, I asked that an hour ago. I totally forgot about it because I chased a different thread.”

Both of us suspect this is not purely a capability issue. The more likely explanation is that OpenAI is managing compute costs by throttling processing per request. That is understandable from a business perspective. But from a user experience standpoint, a model that makes you forget you submitted a request is not a model you are going to keep using for day-to-day work.

GPT-4.5 Instant: Faster, But Inconsistent

The Instant version addresses the speed complaint, but introduces a different one. The responses come back faster, but the consistency is not there.

Sean found that asking the same question at different times produced meaningfully different answers, not just varied phrasing but different conclusions entirely. That kind of inconsistency is a problem if you are using a model to make decisions or to generate content you plan to rely on.

More testing is needed before drawing firm conclusions, but the early read is that Instant trades one problem for another.

Where GPT-4.5 Actually Shines

To be fair, there are use cases where the model delivers.

Deep research is the clearest example. James ran a request that processed for four hours and came back with results he described as extremely thorough. If you are using GPT-4.5 Pro specifically for long-form research tasks and you are not in a hurry, the output quality is genuinely good.

Medical questions are another area where OpenAI has a real edge. James has noticed that when he gives it questions about blood work or medical history, OpenAI returns more thorough responses than Claude and is less restrictive about what it will engage with. That is a meaningful differentiator for anyone using AI to help navigate health information.

And then there is Sora.

Sean has been using Sora 2, OpenAI’s text-to-video tool, regularly and considers it the current leader in that category.

“If you get access to the OpenAI app, you can use it to create a character that is basically you,” Sean explained. “Take a couple of selfies from different angles, read some words on screen, and it creates a version of you that you can put in any scenario.”

The videos include audio, the quality is impressive, and the creative range is wide. For content creators looking to produce B-roll, experimental content, or just something entertaining, Sora 2 is worth the subscription on its own.

The Model Confusion Problem

One consistent frustration, not unique to this release but worth naming, is that OpenAI’s model lineup has become genuinely confusing. GPT-4.2, 4.3, 4.4, 4.5, Instant, Thinking, Pro. Which one is for what?

There is an auto-selection tool that is supposed to route requests to the right model. James has found that it defaults to the cheapest option even when the request clearly warrants something more capable. His workaround is to manually stay on Pro and ignore the auto-selector entirely.

What both of us would like to see is clearer segmentation. Tell people this model is for coding, this one is for research, this one is for document creation. Match the tool to the job in plain language. Some AI companies have been better about this than others, and OpenAI has room to improve here.

What This Means for You

A few practical takeaways from this episode.

If you are on a paid subscription and debating whether to cancel and switch, James’s advice is to hold off. At $20 a month, the constant churn of switching between apps is not worth it. Models take the lead for a week or two, then another company releases something and the order shifts. Stay diversified if you can.

If you are already using OpenAI and want to get the most out of the new models, here is how to think about it. Use GPT-4.5 Pro for deep research where you can afford to wait and the thoroughness matters. Use Sora 2 if you are creating video content. For everything else, especially anything that requires quick turnaround or involves coding, Claude is still the stronger choice right now.

And regardless of which tool you prefer today, keep testing the new releases as they come. The landscape changes fast. A model that underperforms this week may be significantly better next month. Staying current on what each tool can do is part of using this technology effectively.