This week brought big changes to how AI agents help with coding. The most exciting news came from OpenAI, which released a special version of its GPT-5 model just for coding work. This new GPT-5 Codex is much better at understanding what programmers want and can work for hours on tough coding problems.

Replit, a popular coding platform, also made waves with its Agent 3 release. This AI coding helper can now work by itself for up to 200 minutes, compared to just 2 minutes before. It can even test and fix its own code while building apps. Replit grew from making $2.8 million to $150 million in less than a year, showing how much people want AI coding help.

But there's a problem brewing. Companies that make AI coding tools are facing huge costs for running their AI models. The fun days of cheap AI coding (called the "summer of vibe coding") are ending. Many companies now have to charge more money or even think about closing down.

Some good news came from the coding world too. A team built RustGPT, which is a complete AI language model made entirely in the Rust programming language. This shows that developers are creating AI tools using different programming languages, not just the usual ones.

OpenAI also won a big coding competition this week, proving that their AI tools are getting really good at solving hard programming problems. Meanwhile, Amazon is working on something called Kirro IDE, which is a new tool to help programmers write code.

All these changes mean that AI agents are becoming the main helpers for people who write computer programs. These AI helpers can work alone for long periods, understand complex instructions, and even check their own work. However, the high costs might slow down how fast these tools improve.