Anthropic pushed another batch of Claude Code changes on August 29, 2026, framed by the @ClaudeDevs account as "shipping a lot of improvements based on your feedback." Unlike the recent auto mode default rollout or the effort display bug, this update is not one headline change — it is a cluster of startup-latency, install-size, memory, and token-visibility fixes that land together. This post breaks down what actually changed, how to verify it on your machine, and which items matter for multi-agent and loop-heavy workflows.
The one-line action: run claude update, then claude --version.
TL;DR — what shipped and what it means
| Question | Answer |
|---|---|
Why does claude feel faster to launch? | The CLI no longer waits for the sandbox and MCP servers to initialize before you can start typing; claude also skips setup steps it doesn't need for the current command. |
| How much smaller is the Linux download? | The Linux x64 build is 4.5x smaller, down to roughly 75 MB. |
| Does it use less memory? | Native builds use 40-70 MB less memory per session. |
What's new in /cost? | A dedicated prompt-cache line, separating cached context from fresh tokens. |
What's new in /usage? | A per-/loop breakdown by runs and tokens, so you can see which loop is expensive. |
What's new in /tasks? | Each subagent now shows which model and effort level it ran on. |
| What's the Auto mode tab? | A new tab in /permissions to read default auto-mode rules and add your own without editing ~/.claude/settings.json. |
| Remote Control fixes? | Auto-reconnect on server-side disconnect, reliable phone notifications after reconnect, and live view of a foreground subagent's tool calls. |
Faster CLI startup: the sandbox no longer blocks your first keystroke
The most immediately noticeable change is startup latency. Previously, launching claude blocked on two slow initializers — the OS sandbox layer and any configured MCP servers — before the input prompt became usable. On a project with several MCP servers, that could mean a visible pause between running the command and being able to type.
The update decouples those. The sandbox and MCP servers still initialize, but asynchronously — you can start composing your prompt while they come up in the background. Anthropic also notes that running claude now "skips setup it doesn't need," meaning subcommands that don't require the full agent runtime (version checks, config reads) short-circuit the heavier initialization path entirely.
This sits on the same performance track as the idle-triggered garbage-collection fix that cut p99 CPU in half earlier in August — a continuation of the Bun runtime work, not a separate rewrite. None of it requires a settings change.
Smaller download, lower memory
Two related numbers from the same announcement:
- Linux x64 download: 4.5x smaller, ~75 MB. This matters most for CI pipelines and container images that install Claude Code on every run — a smaller binary means faster cold builds and thinner images.
- Native builds: 40-70 MB less memory per session. For multi-session and parallel-agent setups where several Claude Code processes run at once, that saving compounds across every concurrent process.
Verify the memory change yourself: run claude on a task, then check RSS with ps -o rss= -p $(pgrep -f claude | head -1) before and after updating.
More token visibility: /cost, /usage, and /tasks
The second theme is spend transparency — three commands gained breakdowns that were previously opaque.
| Command | Before | After Aug 29, 2026 |
|---|---|---|
/cost | Input, output, and total token counts for the session | Adds a prompt-cache line, so cache reads/writes are visible separately from fresh tokens |
/usage | Aggregate usage against your plan limits | Adds a per-/loop breakdown: runs and tokens for each loop task |
/tasks | List of running and completed subagents | Adds the model and effort level each subagent ran on |
The /usage change is the practically useful one for anyone running loops in Claude Code. A loop that fires every few minutes with a large context can quietly dominate your token budget; until now, /usage gave you no way to attribute consumption to a specific loop. Now you can see "this loop ran 44 times and burned X tokens" and decide whether to widen its interval, trim its prompt, or drop it.
The /tasks addition — model and effort per subagent — closes a visibility gap that made the effort display confusion worse: when you dispatch subagents, you can now confirm what each one actually ran on rather than assuming it inherited the parent's settings. See the model vs effort guide for why that distinction changes output.
Auto mode tab in /permissions
Auto mode — the classifier that screens tool calls instead of prompting you for each one — became the default for Pro, Max, and Team plans on August 14, 2026. Until now, customizing its rules meant hand-editing ~/.claude/settings.json and knowing the schema.
The update adds an Auto mode tab to the /permissions interface. It surfaces the default rule set so you can read what auto mode allows and blocks out of the box, and lets you append your own allow/deny rules through the UI. The underlying storage is still settings.json (see the settings.json reference for the full schema), but you no longer have to open the file to make a routine change.
Remote Control: reconnect instead of exit
Three fixes to Remote Control, the feature that lets you drive a Claude Code session from your phone:
- Server-side disconnect now reconnects. If the session dropped because of something on Anthropic's end, the CLI previously exited. It now attempts to reconnect and resume.
- Notifications survive a reconnect. Permission prompts and new messages reliably appear on your phone after the session reconnects — previously they could be silently lost.
- Live foreground subagent view. You can now watch a foreground subagent's tool calls stream in real time, rather than only seeing the final result.
What this update does not address
The @ClaudeDevs thread drew community replies that the update itself does not resolve — worth naming so you know what is and isn't fixed:
- Desktop app + Remote Control. Several replies asked about connecting Remote Control to the Claude desktop app rather than the mobile web view. Not addressed here; the desktop-side story is the auto-continue / usage-limit behavior and the Aug 28
/resumecross-surface session handoff, both covered separately. - "Claudish / fablish" garbled output. Complaints about occasional garbled or corrupted model output continue in the replies. No fix is claimed in this update.
- Support-ticket response times. Multiple users raised slow support responses. This is an operational issue, not a CLI one, and the update does not speak to it.
- Opus 5.1 routing speculation. Some replies speculated about model routing changes. Nothing in the August 29 announcement confirms or mentions a new model.
Treat the version specifics here as Anthropic's own announcement, delivered via a corporate social account with before/after numbers — the startup and install-size mechanics are easy to verify locally, the exact percentages are not independently audited.
How to verify the update on your machine
# 1. Update
claude update
# 2. Confirm the new build
claude --version
# 3. Time a cold start (should feel snappier; sandbox/MCP load in background)
time claude --version
# 4. Check the new token breakdowns mid-session
/cost # look for the prompt-cache line
/usage # look for the per-loop breakdown
/tasks # look for model + effort columns on subagents
# 5. Open the new permissions tab
/permissions # navigate to the "Auto mode" tab
If /cost still shows no prompt-cache line after updating, you are on an older build — re-run claude update and check claude --version against the release notes.
Related reading
- Claude Code auto mode becomes the default for Pro, Max, and Team
- Claude Code CLI CPU fix: idle-triggered garbage collection
- Claude Code effort 10-100 display bug — not a downgrade
- How to run loops in Claude Code
- Claude Code cross-session messaging and listing agents
- Claude Code Remote Control: drive a session from your phone
- Claude Code /resume now pulls terminal sessions into the desktop app
- Claude Code commands: complete slash command reference
- Claude Code settings.json complete reference
- Claude Code model vs effort: knowing more vs trying harder
Primary source: @ClaudeDevs on X, August 29, 2026 — "We've been shipping a lot of improvements to Claude Code based on your feedback."
Version numbers, download sizes, and memory figures reflect the @ClaudeDevs announcement as published on August 29, 2026. Claude Code ships frequently — run claude --version and check the official release notes before assuming this behavior matches your build.