According to the OpenAI CEO, the just-released Codex-5.3 model uses "Less than half the tokens of 5.2-Codex for same tasks"

That one line already says a lot. There is no assumption anymore that compute or budget is infinite in 2026.

But if you can get better modeling performance while using fewer tokens, that's a win-win for both OpenAI and users.

The 57% on SWE Bench Pro is big if true! But as I mentioned many times in the past, benchmarks can be hit or miss. Will be trying it out in practice in the coming days (my latest fav application of code LLMs is to write native macOS apps for all kinds of productivity tasks).