A Look at OpenAI’s Latest Innovations

Image - The Verge

It’s been almost three years since OpenAI released its flagship innovation, ChatGPT, to the public. Since then, it has made countless improvements and additions to ChatGPT, some of which I’ve covered in The Barefoot Times, including the ChatGPT Plus subscription and an in-depth look at GPT-4, around late 2023. We’re now on GPT-5, the latest model from OpenAI, along with a few other incredible tools that they’ve released since then. This article will go into detail on the newest stuff from OpenAI, explain its use cases, and reflect on its overall helpfulness.

First up: GPT-5. GPT-5 is the latest ChatGPT model from OpenAI and comes with a suite of improvements. This includes the ability to think before it gives you an answer and recognize if a prompt requires some deeper thinking, or if it can give an accurate answer quickly. GPT-5 also comes with GPT-5 Thinking and GPT-5 Pro, which force the model to think longer and improve its ability to work out complex problems. When evaluated on benchmarks like Humanity’s Last Exam and GPQA Diamond, tests that are composed of many advanced questions across a range of subjects, both models performed exceptionally well, with GPT-5 Pro scoring an 89% on the GPQA Diamond and a 42% on Humanity’s Last Exam. Charts below show comparisons of how each model did on each benchmark.

GPT-5 is OpenAI’s most advanced model in terms of knowledge, helpfulness, friendliness, and speed. Open AI combined all of GPT-5’s predecessors into a single model, removing the confusion of the model picker that Open AI previously received negative feedback about. For example, instead of GPT o3 for reasoning and GPT 4o for everything else, now it’s just GPT 5 Thinking and GPT 5.

The next new tool from OpenAI is Agent, a new way that allows ChatGPT to use the web. Agent has access to its own Google Chrome browser, and is trained to complete full tasks exclusively on the web. It does so using its live vision capabilities to recognize what it sees on a page, and its ability to execute clicks and keyboard types to fill out forms and click buttons on your behalf. This means that if you really wanted to, you can now have ChatGPT order DoorDash straight to your door, all on its own. All you need to do is log in to your DoorDash account for the first time, and the first time only, allowing it to remember you for future requests. Agent allows users to almost fully automate very specific tasks, like finding and booking a hotel or listing an item for sale. Below are a few screenshots of Agent in action finding premium economy flights to Japan in early December. You’ll notice how ChatGPT has its own cursor over the live browser view, so you can watch its actions and thinking process in real time.

As Agent is working, another GPT-5 model is standing by and monitoring its work to make sure it stays on track and doesn’t get sidetracked or start to hallucinate. It acts as a guardrail for Agent.

Finally, let’s talk about Deep Research, another new tool that OpenAI recently released and implemented into ChatGPT. This tool uses GPT-5 to search for lots of information over an extended period of time. Deep Research can generate a detailed report on a very specific topic or query, or it can also find the answer to a highly specific question that may not be immediately available on the surface, or to something that you simply don’t have the time to research. Deep Research takes all the time it needs to find information before giving it all to you in its report. This process can take anywhere from 5 minutes to a full hour.

OpenAI has released some super advanced tools that they’ve made available to users, from the latest GPT model with great benchmark scores, to a highly detailed report on a specific question, to even a tool that can book you a hotel in Paris for a weekend (or longer!).

Next
Next

Magnolia Bakery: Delicious but Pricey