Anthropic pushed a focused but practically significant update to Claude Code on March 17, raising the token output ceiling across its current generation of models and automatically migrating the remaining Sonnet 4.5 users to Sonnet 4.6.
The changes to output limits are the most immediately useful part of the update. The default maximum output for Claude Opus 4.6 in Claude Code has increased to 64k tokens, and the upper bound for both Opus 4.6 and Sonnet 4.6 now reaches 128k tokens. For developers working on long codebases, large refactoring tasks, or multi-file generation, the practical effect is fewer interruptions and more complete outputs per session. Previously, hitting output limits mid-task required restarting or breaking work into smaller chunks. The higher ceiling reduces that friction.
The automatic migration from Sonnet 4.5 to Sonnet 4.6 for Pro, Max, and Team Premium users is the other substantive change. Sonnet 4.6, released on February 17, delivers improved agentic search performance and consumes fewer tokens than its predecessor while supporting a 1 million token context window. Users who had not switched manually are now on the newer model by default, which means better performance on the tasks where Claude Code is most commonly used: coding, multi-step agent workflows, and computer use.
Alongside the limit increases and model migration, the update includes a number of fixes with direct impact on developer workflows. The /copy command now accepts an index, so users can retrieve specific past assistant responses rather than only the most recent one. A bug that caused “Always Allow” rules on compound bash commands to save a single rule for the full command string rather than per-subcommand has been resolved. This had created dead rules and repeated permission prompts for anyone using chained commands like cd src && npm test.
The auto-updater received a fix for a memory issue where overlapping binary downloads could accumulate tens of gigabytes if the slash-command overlay was opened and closed repeatedly. A race condition in –resume that silently truncated recent conversation history has also been patched, along with a fix for PreToolUse hooks that could bypass deny permission rules including enterprise managed settings.
This update does not represent a major product milestone in the way that the release did, but the compounding effect of consistent engineering improvements to Claude Code is part of what has made it the leading AI coding assistant heading into Q2 2026. The token limit increases in particular are the kind of change that developers who use the tool daily will notice without needing a changelog to tell them something changed.
