Claude Code's 29 August update: faster startup, a much smaller Linux download, and better token visibility
Claude Code now starts before the sandbox loads, the Linux binary shrinks from ~340 MB to ~75 MB, and /cost, /usage, and /tasks get per-command token detail.
The 29 August batch update for Claude Code is the kind of release that doesn’t make headlines but makes your day noticeably better. Three things stand out: the CLI prompt appears before the sandbox and MCP servers have finished loading, the Linux binary is roughly four and a half times smaller, and the /cost, /usage, and /tasks commands now show you a proper breakdown of where your tokens went. None of these are flashy, but all of them are immediately useful.
The prompt appears faster, even if the sandbox isn’t ready yet
Previously, Claude Code waited for the sandbox and any MCP servers to finish initialising before showing you the prompt. That wait is now gone. Sandbox and MCP bring-up happen in parallel with the UI loading, so you see the prompt straight away. Bare launches also skip subcommand registration, and workflow discovery, settings, and trust-store work is lighter overall.
What does this mean for you? If you open Claude Code frequently, or if your workflow involves spinning it up inside scripts or containers, you’ll notice the improvement immediately. The tool feels more responsive, and you don’t have to sit watching a blank screen waiting for MCP connections to settle before you can type.
The Linux download is now ~75 MB, down from ~340 MB
On Linux x64, the native binary has been recompressed with zstd, bringing the download from roughly 340 MB to roughly 75 MB. The native binary itself is also about 2 MB smaller, thanks to more compact storage of bundled skill and prompt text.
If you’re on a fast connection this might feel like a minor convenience. If you’re updating Claude Code inside a CI pipeline, a remote server, or a container image build, that 4.5x reduction is material. Faster downloads, smaller images, and less bandwidth cost all follow from a single change. Anthropic recommends updating immediately if you run Claude Code inside rootless containers or user namespaces, but the smaller footprint is worthwhile for everyone.
To get the update, run claude update and then claude --version to confirm.
Token visibility: /usage, /cost, and /tasks now show you the detail
This is probably the most practically useful cluster of changes in the batch, particularly if you’ve been running multi-agent or loop-heavy workflows.
/usage now includes a loops breakdown. For each loop task, you can see the run count, total tokens consumed, tokens per run, and the timestamp of the last run. This matters because a loop that fires every few minutes with a large context can quietly eat through your token budget. Before this update, you’d see an aggregate figure and have to guess which loop was responsible. Now you can see “this loop ran 47 times and consumed X tokens per run” and make an informed decision about whether to widen its interval, trim its prompt, or remove it entirely.
/tasks now shows the model and effort level each subagent ran on. When you dispatch subagents, there was previously no easy way to confirm what model each one actually used. The assumption was often that subagents inherited the parent’s settings, which wasn’t always true. The model and effort level are now visible both in /tasks and in the agent detail dialogs.
/cost now includes a per-model and cache-hit breakdown for subscription users. This is a straightforward but welcome addition. If you’re using multiple models across a session, or if prompt caching is saving you tokens, you can now see exactly where the cost is coming from. An interactive /release-notes version picker has also been added in this batch.
This timing is deliberate. Auto mode, which screens tool calls automatically rather than prompting you for each one, became the default for Pro, Max, and Team plans on 14 August 2026. More automation means more token consumption happening in the background, so visibility into that consumption has become more important. These three commands now give you the tools to understand what’s actually running and what it’s costing.
Other additions worth knowing about
A few smaller but useful features shipped in the same batch.
A modelPicker setting lets you define a custom, ordered list of models for the /model picker, including Vertex and Bedrock model IDs. You can append to the built-in lineup or replace it entirely with your own.
Two new settings, promptCacheTtl and subagentPromptCacheTtl, give API-key and cloud-provider users control over prompt cache lifetimes. The intended pattern is a 1-hour cache on the main conversation with a shorter 5-minute cache for subagents, but the settings are separate so you can configure them independently.
A modelPricing managed setting lets organisations specify contracted per-model rates and a discount multiplier, so /cost, the status line, and telemetry figures reflect actual contracted pricing rather than list price.
Finally, a keyless sign-in option has been added under /login, via Anthropic Console. Organisations that don’t permit API keys can now authenticate without one.
Bug fixes that matter
Two fixes in this batch are worth calling out. Remote MCP servers in non-interactive (-p) and SDK sessions would previously never recover after a dropped connection. They now reconnect automatically or report as failed, which is significant if you’re running Claude Code in headless environments where you won’t notice a silent failure. A separate fix addresses fast arrow-key and Enter sequences acting on the wrong row during history search and several other UI contexts, including /config, /mcp, /skills, background tasks, and /model.
Run claude update to get everything, and check claude --version to confirm you’re on the 29 August build.