Aug-Oct 2025: few big bang announcements, more incremental developments
There has been a lull in big bang announcements for AI dev tools over the last couple of months. However, we have seen a variety of small, incremental developments:
-
Claude has released a new version of its model, Sonnet 4.5. It shows incremental improvements on benchmarks but notably, Anthropic emphasises its ability to work on a task much longer, as well as its higher rate of success in computer use. Claude Code has also acquired some new features like checkpoints.
-
Zed has created a mechanism for integrating agents into its editor via the Agent Client Protocol (ACP). This means that Zed provides the UI and the agent works behind the scenes. They’ve added beta support for Claude Code (usable but buggy). Subsequently, JetBrains has started to adopt ACP in their IDEs.
-
A couple of months ago, z.ai, one of the large foundation model developers from China (not to be confused with Zed the editor), released a new generation of its GLM models.
-
Google published a technical paper on the energy and water use of its Gemini model.
-
For companies, the coming support for MCP controls in GitHub Copilot should be welcome news as malicious MCP servers are an obvious attack vector.
-
Apple’s XCode 26 has been released with a built-in AI agent. The agent can be set up with a ChatGPT, Claude or another model (including local models).
-
I came across a couple of interesting examples of substantial results with mostly AI-generated code: unobtrusive update UI in a Swift application and modernising a 25-year old Linux kernel driver.