Claude Code shipped versions 2.1.311 (26 October 2026) and 2.1.312 (27 October 2026). Most of the notes are bug fixes, with 56 and 22 listed. A few changes are worth a non-developer’s attention, because they show how AI coding tools are being tuned on cost and control.
What the release notes say
- Typing
ultrathinknow runs that turn at xhigh effort on several current models through Anthropic’s API. The user chooses when to spend more effort on a hard problem. - Cloud routines, the scheduled tasks that run without anyone at a keyboard, get an effort-level picker.
- Automatic model switches after a flagged message now also cover unrecognised model names. They can be turned off with a setting.
- On Bedrock, Vertex, Foundry and Mantle, unrecognised model names get a shorter system prompt, unless the full prompt is set explicitly.
- The “5-hour window” is now called the “5-hour limit”.
The pattern
None of these is a headline feature. Together they point at the same thing: effort, model choice and limits are now settings that someone sets. A routine that runs overnight at maximum effort costs more than one that doesn’t. A tool that silently switches model after a flagged message may behave differently on the same task. The tool is no longer one fixed thing you buy.
So what
When an agency says it builds with AI tools, the useful questions are which settings it uses by default, who decides when to turn effort up, and what happens when a limit is reached mid-project. Those answers affect timeline and cost more than the brand of tool. Our AI-assisted development approach keeps a named engineer reviewing the output regardless of effort level. To discuss a build, start a project.