I went through the benchmarks ;
Opus explores, Codex executes. Pick based on how you work. The timing is hilarious though. Same day, same hour basically. The AI wars are getting spicy
- GPT-5.3 absolutely cooked on Terminal-Bench. 77.3% vs 65.4%, not even close. If you're running agents that need to hammer out CLI commands all day, this is the one. Speed is legit too — 25% faster and burns half the tokens for the same work.
- Opus 4.6?? Near-human on OSWorld. 72.7% when humans score ~72%. GPT-5.3 is at 64.7% there. For anything where the model needs to actually think its way through a messy problem on a computer, Opus pulls ahead.
- The 1M context thing is lowkey the biggest deal nobody's paying attention to. Opus scores 76% on needle-in-a-haystack at 1M tokens. That's not "we accept 1M and forget 90% of it" — it's actually tracking stuff across the whole input.
- Every.to ran independent tests and found something that matches what I've been feeling too: Opus is better when you throw a vague problem at it and let it cook. GPT-5.3 is better when you write a tight spec and want precise execution.
Opus explores, Codex executes. Pick based on how you work. The timing is hilarious though. Same day, same hour basically. The AI wars are getting spicy