It’s time for another Counter-Strike benchmark: GPT 5.3 Codex and Claude 4.6 Opus:

Which reasoning effort are you using for Codex?

TL:DR 1. Both models beat all previous generations 2. Codex was 2x faster than Claude 3. Claude beat Codex in just about every prompt Want to play yourself? Codex: Claude: Let's dive into the details:

1. These models are definitely a leap over the previous ones. They made much nicer maps. They made better looking guns. They made very realistic humans. And they made very few coding errors. Just compare the humans:

Codex was a lot faster, but it could be due to the harness. We noticed Claude did a lot more planning. In the end, Claude ended up making better decisions for just about every prompt but one (the last one was a tie). Here's an overview for each prompt:

At this point, neither model had issues with coding or backend. The frontier now seems to be about physics. For example, Claude generated maps where players could end up stuck. Here's Claude's "inferno valley" and "nuke zone" produced 4-wall obstacles in the center:

Codex also had trouble with direction. The enemy's "point of view" was coming out from the back of their had, rather than the front. With both models you could shoot through obstacles.

With either one, it was fun to build and play! One reason they went so far: they could use @instant_db for the multiplayer. Try it yourself. Start with: `npx create-instant-app` and say: "Build me multiplayer Counter-Strike with Instant presence and topics"

cc @Suhail you may find this interesting!

the physics frontier observation is spot on. coding errors are basically solved now but spatial reasoning is where these models still struggle. curious how much longer until physics simulations become trivial for them too.

🔥🔥🔥

How does this test work? Do the models go this fast while looking at the image or just the data as a text?

The main observation is that Counter strike can't be build in one prompt yet !