See at a glance how much your always-on /loop is burning──Claude Code v2.1.243 puts a gauge on "runaway costs" and slims the binary from 340MB to 75MB

Anthropic shipped v2.1.243 of Claude Code on August 25. It tackles the "invisible cost" and "weight" of running agents unattended, making it one of the bigger updates even for this summer. The same day also brought v2.1.245, which fixes a startup crash on Linux. Here we focus on the changes that matter for day-to-day operations.

"Which loop is eating tokens" now shows up as a number

The headliner is the new "Loops" breakdown added to /usage. If you run /loop for scheduled jobs or automation, you could previously only see which tasks were consuming tokens as a lump sum. This update lets you break it down loop by loop.

Four items are displayed:

• Run count: how many times that loop ran
• Total tokens: cumulative consumption
• Tokens per run: the average weight of each execution
• Last run: when it most recently ran

You can now pinpoint, on the spot, a loop with an outsized "tokens per run" or a task you forgot to stop that keeps running. In unattended operation, it's all too easy to only notice once the bill has ballooned; this is a gauge for nipping that in the bud early.

From a 340MB binary to 75MB──a diet for startup and size

Another standout change is the reduction in distribution size and memory. The native binary is now compressed with zstd, shrinking it from roughly 340MB to 75MB on Linux x64. Downloads and auto-updates get lighter, which pays off in CI and when distributing to many machines.

The memory reduction comes from "on-demand loading" — instead of keeping the entire bundle resident, only the parts you need are loaded. Each session gets about 40–70MB lighter, so the more agents you run in parallel, the more the savings add up. Startup is also faster: launching the sandbox and MCP no longer holds up the first render, and a plain launch now skips registering subcommands.

Refinements to model selection, caching, and login

The finer details of operation got attention too. Since these changes come with specific counts, we lay them out item by item.

modelPicker setting

You can now curate the /model options yourself as an ordered, labeled list. The aim is to narrow the models a team uses, cutting down on hesitation and mis-selection.

Prompt cache retention time

promptCacheTtl and subagentPromptCacheTtl were added. Users on API keys or via cloud providers can set separate cache lifetimes — for example, one hour for the main conversation and five minutes for subagents. This curbs the cost of re-sending prompts when the cache expires during long tasks.

Keyless console sign-in

You can now choose "Sign in with your Anthropic Console account" from /login (recommended). Because you can log in without keeping an API key on hand, it reduces key management overhead and the risk of leaks.

The same-day v2.1.245 is a hotfix for a startup crash

v2.1.245 fixes a bug where Claude Code crashes on startup on some Linux distributions that adopt glibc 2.44. Affected systems include Arch Linux, CachyOS, and Fedora Rawhide. If you're on a bleeding-edge environment, it's safest to update to at least this version.

Cost visibility is welcome — but it's also a caution against "over-running"

This update centers on real benefits for those running Claude Code unattended and in parallel. The per-loop cost breakdown keeps usage healthy, and the cuts to size and memory make it easier to run even as you scale up machines. The lower the barrier to adoption, the more running agents around the clock will spread.

At the same time, visibility is a feature that presupposes overuse. The more you keep loops running, the more token consumption, compute, and by extension cost and power load pile up. The /usage breakdown is meant to be used for identifying which loops to trim. Rather than letting convenience run everything nonstop, run only the processing you actually need — only paired with that kind of operational discipline does this new gauge start to bite.

References: Claude Code changelog (official) / anthropics/claude-code CHANGELOG.md / Releasebot: Claude Code updates