GPT-5 is here, and it’s no surprise

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

As weird as an AI funeral is, I realized that I probably spend more time talking to ChatGPT than some of the humans whose funerals I would attend. And honestly, not all of them are as nice as ChatGPT.

In today’s issue…

• GPT-5 underwhelms its insane expectations

• Google’s fully-autonomous coding agent

• A local, open-source AI app builder

• How AI is reinventing software developers

• A deep dive into the top agent communication protocols

…and more

🔍 SPOTLIGHT

More than two years after the release of GPT-4 changed the world by launching the AI agent field, GPT-5 is out—and this time, the world doesn’t feel that much different.

Last Thursday, OpenAI rolled out their much-hyped new flagship large language model (LLM) to ChatGPT users, replacing all of their previous models with a single GPT-5 option (well, all but one—more on that in a moment). The older models are still available to agent builders through the API, but with the GPT-5 series outperforming all of them and at a lower cost, there’s little reason not to switch over.

When I say “outperforming all of them”, though, the magnitude matters. If, as many people predicted, GPT-5 had been as much of a leap over GPT-4 as GPT-4 was over GPT-3, we might have been having serious conversations about whether artificial superintelligence had arrived. I treated this as a very real possibility in my June 23rd Spotlight on the GPT-5 rumors when I explored the question of whether GPT-5 would be evolutionary or revolutionary.

So, which was it? Definitely evolutionary.

GPT-5 outperforms o3, OpenAI’s previous strong contender for “best model in the world” across all the major benchmarks—but not by the huge margins that GPT-4 had over its predecessors. From a “vibes” standpoint, the near-universal consensus was that the release was pretty anticlimactic. Enough ChatGPT users actually preferred the much older GPT-4o that they managed to bully OpenAI into restoring it as an option.

GPT-5 is still a great model, probably the best in the world. It may not be as “warm” as 4o in some people’s minds, but its unquestionably smarter, which is far more relevant for agent builders. Perhaps most importantly, it’s affordable—priced much lower than o3 on input tokens ($1.25 vs. $2), and only slightly higher ($10 vs. $8) for output. Its smaller versions, GPT-5 Mini and Nano, are correspondingly cheaper. Personally, all of my agents are being switched to GPT-5 immediately, and if yours aren’t, you’re almost certainly throwing away easy performance gains and cost savings.

But if you were hoping for a friendly version of Skynet, you may have to wait for GPT-6.