September brought many releases across Anthropic, Google, xAI and OpenAI. Most of them are mainly aimed at developers. Let's take a look at what the news brought this time.
Claude 4.5 Sonnet
Anthropic introduced Sonnet 4.5, which significantly advances autonomy and reliability for coding and agent scenarios. Although it is more of a mid-range model in terms of performance, it is stronger in some tasks than, for example, Claude 4 Opus. In internal and external demos, it managed to run autonomously for tens of hours (jumping from ~7 h to ~30 h), bringing "context editing" and a new memory tool in the API for longer agent runs without losing important information. At the same time, Anthropic released the Claude Code IDE Extension.
What this means in practice: fewer hallucinations, better instruction following, more stable long tasks (e.g. refactoring of larger parts of the code) and more comfortable working directly in the editor thanks to the official plugins.
Useful resources:
Gemini 2.5 Flash & Flash-Lite
Google has released a preview of two models. They are interesting mainly because of their speed. They can handle up to 880 tokens per second and that's a huge improvement, for comparison, for example Claude 4.5 Sonnet can handle 70 tokens in the same time.The great thing is that the price is still the same even with this performance, so the model is ideal to use for products that depend on fast LLM result or agile UX.
Useful resources:
Year 4 Fast
xAI introduced Grok 4 Fast, a reasoner optimized for efficiency: it achieves performance comparable to high-end models in internal metrics, while typically using ~40% fewer "thinking" tokens. This translates into lower costs while maintaining quality. It will be great to use it especially where you are addressing cost scaling for reasoning.
Useful resources:
GPT-5 Codex
OpenAI pushes Codex, a cloud-based coding agent, and adds GPT-5-Codex as a specialized model. In addition, Codex CLI and IDE extension for deeper workflow integration. OpenAI's effort to unify the Codex family across tools is clearly shown here. The goal is better integration and increased usability without unnecessary context switching.
Useful resources:
Gemini 3 coming soon
We may see Gemini 3 as early as October, but not before early December. Google hasn't come up with any official release date yet. We can expect it mainly by Google's steps of the last years, when it always released new versions during Q4. (Gemini 1.0 in December 2023 and Gemini 2.0 in December 2024) But a limited preview could appear as early as October.
In general, we can observethat companies are focusing more and more on the infrastructure around developers and on integration into developer environments. In the meantime, however, the price of models and the capabilities of foundation models remain more or less stable. The key question is how to ensure that users ask the right questions and models prove their full value.