Hello and welcome to the 16th issue of AI Powered Engineering, and it’s an exciting one because 2026 has been the most jam-packed year of releases we’ve ever seen. Yesterday was a great example of the pace we’re seeing with releases this year as Anthropic dropped Opus 4.6, and then 29 minutes later - boom, there was GPT 5.3 Codex.

Here’s Anthropic’s tweet announcing Opus 4.6:

Less than half an hour later (by a minute) Sam dropped this:

I’m so excited about these two releases, I decided I’m just going to dedicate this entire issue to them. That being said, I have a lot of other topics I originally wanted to cover in this issue, so I might just have to put out issue #17 earlier than usual, there’s just too much good stuff. Okay, but I know what you want, time to get to the good stuff, let’s rock 🤘

After spending the afternoon with both models, I think I have a solid high-level summary of each, and, I can also say pretty clearly now, one isn’t better than the other, they are pretty different.

With Opus 4.6, the focus is deploying autonomous teams of agents that go off and do your bidding. Anthropic has a great example of how you would use this new functionality with a sample prompt on their site. I think this does a good job of illustrating how you might think about building teams of agents.

What’s so interesting here is that if you think more about the agentic team that you’re creating with Opus 4.6, you can give them independent roles, so they can actually each explore the problem, without waiting on each other.

This is the killer feature in Opus 4.6 if you look beyond the benchmarks. It’s designed for people who wanted to use AI Powered Engineering to spin up teams of agents that work away for them.

Now with GPT 5.3 Codex, the focus is a little different, and in some ways the opposite. I think the best way I’ve heard it described is as OpenAI talks about it themselves, an interactive collaborator. GPT 5.3 is like that insanely cracked engineer that you’d love to pair with, and now you can.

Rather than spinning up a bunch of agents, GPT 5.3 Codex is for people who want to really go deep with one agent. The killer feature here in many ways is the ability to steer the agent mid-execution.

Okay, I’ll stop there because I really wanted this first section to be short and sweet. You should now have a pretty clear picture of how Opus 4.6 and GPT 5.3 Codex are different. And, the gears might be turning for you now as you think more about which style you prefer…or if maybe you might want to combine the two?

A few after Opus 4.6 launched, I was seeing quite a few confused people on X, that either couldn’t figure out if they were running the newest version of Opus, or couldn’t get agent teams to work.

So I jotted down some notes in Obsidian, and thought I may as well share them here.

Opus 4.6, initial todos:

Potential additional configs:

The biggest issue I saw people running into was just not updating Claude Code. So follow those first two steps above and make sure you’re not still on a 1.x version.

Next, agent teams isn’t turned on by default, you need to go into settings.json and set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS equal to one. And if you want the model to default to opus, add that to your settings.json too.

Also, if you’re using the API, don’t miss the updates Anthropic made to adaptive thinking:

One note, if you do crank it up to max, don’t be shocked if your bill goes through the roof. This is a super cool feature, but one you definitely want to be careful with.

Now that OpenAI has the Codex app, you don’t really have to do much setup to start using all the features in GPT 5.3 Codex. That being said, I still saw some people, sharing picts and videos on X, thinking they were using GPT 5.3 Codex, when they were still using GPT 5.2.

So make sure that you’re selecting GPT-5.3-Codex from the dropdown menu in the Codex App. And yes, I know that Codex isn’t available for Windows yet, and I do see this as a real drawback. While it might be unpopular to say this, what the heck, I’ll just say it.

If you’re building software in 2026, you probably should be doing it on a Mac. If you aren’t, well then, you’re missing out on getting to use the latest-and-greatest tooling.

Also, just like Anthropic has effort tuning, ChatGPT has pretty much the same thing, it’s just to the right of the model selector, and allows you to go between Low, Medium High, and Extra High reasoning modes.

Okay, well, I can’t write about these models without sharing the benchmarks. Since Anthropic went first, they were able to publish benchmarks comparing to GPT 5.2, which essentially was pretty much instantly outdated with the release of 5.3. But that’s just how it goes sometimes.

Now with that being said, here’s the benchmarks for each:

Opus 4.6 Benchmarks

GPT 5.3 Codex Benchmarks

I think it’s safe to say, opinions online are pretty split on these two models. Here’s some tweets about each from people who I’d say know their stuff.

Now it’s your turn, if you haven’t yet - get in there, play with both new models, try everything you can think of, and figure out which one works the way you want to work. For me, I’m actually starting to explore different ways to use them together, and already discovering some pretty interesting stuff.

Okay, that’s a wrap, see you next time, whenever that is!