September AI news

September brought a wave of releases from Anthropic, Google, xAI, and OpenAI. Most of them are aimed mainly at developers. Let's take a look at what landed this time.

Anthropic introduced Sonnet 4.5. It clearly advances autonomy and reliability for coding and agent scenarios. In raw performance it's more of a mid-range model, but on some tasks it beats even Claude 4 Opus. In internal and external demos it ran on its own for tens of hours (a jump from roughly 7 to roughly 30 hours). It brings "context editing" and a new memory tool in the API, so agents run longer without losing important information. Alongside it, Anthropic released the Claude Code IDE Extension.

What this means in practice: fewer hallucinations, better instruction following, more stable long tasks (such as refactoring larger parts of the code), and more comfortable work right in the editor thanks to the official plugins.

Useful resources: About Claude

Google released a preview of two models. They're interesting mainly for their speed. They handle up to 880 tokens per second, which is a huge jump. For comparison, Claude 4.5 Sonnet handles 70 in the same time. The great part is that the price stays the same at this performance, so the model fits products that depend on a fast LLM result or a snappy UX.

Useful resources: Artificial Analysis Google

xAI introduced Grok 4 Fast, a reasoner tuned for efficiency. In internal metrics it matches high-end models while typically using about 40% fewer "thinking" tokens. That translates into lower costs while quality holds. It's a strong fit wherever you need to keep reasoning costs in check as you scale.

Useful resources: Artificial Analysis Grok

OpenAI pushes Codex, its cloud-based coding agent, and adds GPT-5-Codex as a specialized model. On top of that come the Codex CLI and an IDE extension for deeper workflow integration. OpenAI's push to unify the Codex family across tools shows clearly here. The goal is better integration and higher usability without needless context switching.

Useful resources: OpenAI Upgrades

We may see Gemini 3 as early as October, and no later than early December. Google hasn't announced an official release date yet. We base this mainly on Google's pattern in recent years, releasing new versions during Q4 (Gemini 1.0 in December 2023 and Gemini 2.0 in December 2024). A limited preview, though, could show up as early as October.

In general, we can see that companies are focusing more and more on the infrastructure around developers and on integration into developer environments. Meanwhile, the price of models and the capabilities of foundation models stay more or less stable. The key question is how to make sure users ask the right questions, so the models can show their full value.