Claude Code 2.1.243 Ships Loops Breakdown in /usage, Custom Model Pickers, and a 75 MB Binary

Two releases landed on August 25 in the Claude Code changelog. Version 2.1.245 patched a startup crash on Linux systems running glibc 2.44 (Arch, CachyOS, Fedora Rawhide). The bigger one, 2.1.243, is a sweeping update that covers observability, model selection, pricing, authentication, and a long list of reliability fixes.

Loops finally show up in /usage

If you run recurring /loop or /goal tasks, the /usage command now breaks them out separately. Each loop gets its own row: run count, total tokens, tokens per run, and a timestamp for the last run. It’s the kind of data that helps you spot a loop that’s running too often or burning more context than expected.

Curate your own model picker

A new modelPicker setting lets you define exactly which models appear in /model — and in what order. You can use any model ID, including Vertex and Bedrock IDs, and you can add a human-readable label for each. The list either appends to the built-in lineup or replaces it entirely. For teams on managed instances that want to limit engineers to a specific set of models, this is the knob they’ve been missing.

Cache settings for long sessions

Two new settings control prompt caching behavior. promptCacheTtl keeps the 1-hour cache active on the main conversation (useful for long-running sessions where the first few turns rarely change). subagentPromptCacheTtl stays at 5 minutes, since subagents churn through more varied context. Organizations can also use a new modelPricing managed setting to apply their contracted per-model rates and discount multipliers to /cost, the status line, and telemetry — so cost estimates reflect what they actually pay.

Logging in no longer requires creating an API key. A “Sign in with your Console account” option under /login lets users authenticate directly. Organizations that block API key creation can use this path instead.

Sonnet 5 price cut baked in

The update reflects a pricing change: Sonnet 5 is now listed at $2 per million input tokens and $10 per million output tokens. That’s down from the previous $3/$15 standard list price. The /cost command and status line both update automatically once you’re on 2.1.243.

Smaller and faster

The binary is now compressed with zstd. On Linux x64, that takes it from roughly 340 MB to about 75 MB. Each session also uses 40 to 70 MB less memory than before, with further reductions in peak memory for long conversations. Startup is faster too: the sandbox and MCP servers no longer block the first frame from rendering, and bare launches skip subcommand registration.

Subagent visibility in /tasks

The /tasks command and agent detail dialogs now show which model and effort level each subagent ran on. When you’re debugging an agent pipeline, that’s useful context that previously wasn’t surfaced.

Bug fixes worth calling out

The list of fixes in 2.1.243 is long, but a few stand out:

Remote MCP servers in non-interactive (-p) and SDK sessions now reconnect automatically after a dropped connection.
MCP OAuth sign-in from the desktop app no longer fails with “Invalid redirect URI” on servers that support client ID metadata documents (Linear users hit this).
Auto mode no longer gets stuck as unavailable after a temporary server-side disable.
Sessions that got no response from the API for 10+ minutes now time out properly after about 3 minutes instead of going silent indefinitely.
/resume now loads more sessions as you scroll, lifting a hard cap that stopped it at 50.
VS Code sessions now open in auto mode instead of defaulting back to manual mode.
claude --teleport no longer exits on uncommitted changes — it offers to stash them first.

The full changelog is at code.claude.com/docs/en/changelog.