Claude Code v2.1.77 Update — 128K Token Output, Task Resume 45% Faster
Claude Code v2.1.77 has been released. Opus 4.6 model output expands 4x to 128K tokens, task resume speed improves by 45%, and over 20 bugs are fixed including a memory leak.
The AI coding assistant Claude Code has been upgraded to v2.1.77. The most notable change is how much code the AI can generate in a single response. The Opus 4.6 model's default output has increased to 64,000 tokens, with a maximum of 128,000 tokens (128K). The frequent issue of AI cutting off midway when asked to write long code is now significantly reduced.
Task resume speed has also improved by up to 45%, and Claude Code launches 60 milliseconds faster on first startup on macOS. A total of over 20 bugs have been fixed, including one security vulnerability and a memory leak bug.
Claude Code 128K Output — Generate an Entire Novel's Worth of Code at Once
A token is the unit AI uses to process text. In English, one token is roughly three to four characters. 128,000 tokens is approximately 90,000–100,000 words — close to the length of a full novel.
In previous versions, asking the AI to write an entire long code file would frequently result in it being cut off midway. With this update, the default output limit is set to 64,000 tokens, expandable up to 128,000 tokens when needed.
In practice, this means being able to generate multiple files of hundreds of lines in one go, or completing a complex feature from start to finish in a single response. This will be especially impactful for those doing vibe coding — instructing AI to build apps using natural language.
Task Resume Speed Up 45% and Memory Savings
Claude Code has a --resume feature that lets you continue from a previous conversation. This allows you to pick up where you left off on a project you were working on the day before.
This update improves the loading speed of this feature by up to 45% and reduces memory usage by 100–150MB. The improvement is especially noticeable in long sessions. What previously took several seconds to resume is now noticeably faster.
On macOS, authentication credentials from Keychain (the password store) are now read in parallel, cutting startup time by approximately 60 milliseconds. A small number, but one that adds up every time you launch Claude Code.
Auto-Update Memory Leak Bug Fixed — Dozens of GBs Reclaimed
The most severe bug in this release was the auto-update feature repeatedly downloading the same file, consuming tens of gigabytes of memory. If you experienced your computer slowing down after leaving Claude Code running for a long time, this bug was likely the cause.
Here is a summary of other key bugs that were fixed:
Security Fix — A vulnerability in the PreToolUse hook (the validation step that runs before AI uses a tool) where an "allow" response could bypass "block" rules has been patched. This could have allowed security policies set by administrators in enterprise environments to be bypassed.
Long Session Memory Leak — A bug where progress messages kept accumulating in memory during extended use has been fixed. Sessions now remain stable over long periods of use.
Conversation History Truncation — A bug where recent conversation content was silently truncated when resuming with --resume has been fixed. Previously, some context could be lost when resuming a session.
CJK (Korean/Chinese/Japanese) Rendering Bug Fixed
A bug where CJK characters such as Korean, Chinese, and Japanese were overflowing into adjacent UI elements on screen has been fixed. This is great news for anyone who writes code comments in Korean or gives instructions to the AI in Korean and experienced display issues.
There are also many additional terminal-related fixes:
- Fixed double-opening of links when using Cmd+Click in VS Code and Cursor
- Fixed clipboard copy failure in tmux (terminal multiplexer) — added a toast notification to confirm copy success or failure
- Backspace and Delete keys now work correctly in vim mode
- Fixed session conflict when using iTerm2 with tmux remote connections
- Fixed arrow keys changing tabs in the settings dialog
New Claude Code Commands — /copy N and /branch
Usability has also been improved:
/copy N command — Previously only the most recent AI response could be copied. Now you can type /copy 3 to copy the response from 3 turns ago.
/fork → /branch rename — The command for branching a conversation has been renamed to /branch. The existing /fork command continues to work as well.
Automatic plan naming — When the AI proposes a plan and the user accepts it, the session is automatically named based on the plan content. This makes it much easier to find a past session later.
Background task safeguard — Background commands are automatically terminated if their output exceeds 5GB. This prevents runaway logging commands from accidentally filling up your disk.
How to Install or Update Claude Code
If Claude Code is already installed, it will update automatically. To update manually, run the following command in your terminal:
claude updateIf you have never used Claude Code before, you can install it with the following commands:
# macOS, Linux
curl -fsSL https://claude.ai/install.sh | bash
# After installation, run from your project folder
cd your-project
claudeA Claude paid subscription or Anthropic Console account is required to use Claude Code. If you are new to Claude Code, the Getting Started with Claude Code guide walks you through installation and first use step by step.
To learn more about AI coding and vibe coding, check out our free learning guide.
Related Content — Get Started with AI Coding | Free Learning Guide
Stay updated on AI news
Simple explanations of the latest AI developments