⚡️⚡️⚡️ OpenAI has released GPT-5.3-Codex
A new coding model with better metrics:
- 57% SWE-Bench Pro (real software engineering tasks in 4 languages)
- 77% TerminalBench 2.0 (terminal skills)
- 64% OSWorld
What's new:
- You can control the agent while it's working, without losing context - just like with a colleague
- 2 times fewer tokens than 5.2-Codex on the same tasks
- 25% faster per token
- Good computer use
Interesting fact: GPT-5.3-Codex is the first model that participated in its own creation. The team used earlier versions for debugging training, deploying, and diagnosing tests.
🔗 openai.com/index/introducing-gpt-5-3-codex
