Almost all major AI coding projects have already added the Gemini 3 Pro model. Antigravity itself often throws connection errors and works very slowly.

GPT 5.1 Codex
https://openai.com/index/gpt-5-1-codex-max/
After updating the regular model to GPT 5.1, its Codex variant was also updated. To respond to Google, they also announced GPT-5.1-Codex-Max. Everything, of course, got even better. And it will be able to work even longer.

Compaction technology allows the model to work with millions of tokens without losing important information. In practice, this opens up possibilities for refactoring entire projects and autonomous work for many hours straight.
The model is already available in the Codex tool, and API access will appear soon.

https://news.ycombinator.com/item?id=45982649
In the discussion, people share success stories. For example, one developer told how Codex completely rewrote key parts of his hobby project (a flight simulator) in 45 minutes, performing complex refactoring that would have taken a human much longer.

Users also compare the new product with Google Gemini and, while noting certain advantages of the latter, most agree that for code generation, Codex still provides consistently better results, "hallucinates" less often, and integrates better into existing code.

An interesting conversation about the "folk magic" of prompt engineering: people add special "canary words" to their instructions (e.g., asking the model to always address them as "Mr. Tinkleberry") to check if the AI is still carefully following the initial instructions in long sessions.

#newllmmodel #codex