Claude Code 2.1.243 lets API-key users keep a 1-hour prompt cache
The same release adds a curated /model picker, contract-rate /cost, keyless Console sign-in, and self-healing remote MCP; a Linux glibc-2.44 crash fix follows.
Copy markdown
1-hour prompt cache, now for API-key sessions
The new `promptCacheTtl` setting keeps a 1-hour prompt cache on your main conversation while subagents stay at the 5-minute default via `subagentPromptCacheTtl`. If your long interactive sessions were silently re-caching context every 5 minutes, this is a direct cut to the token bill on API-key and cloud-provider accounts.
Build your own /model list
`modelPicker` lets you curate the `/model` menu with an ordered, labeled list of any model IDs — Vertex and Bedrock spellings included — appended to or replacing the built-in lineup. Teams pinned to specific models stop scrolling past ones they can't actually call.
/cost that matches your contract
A `modelPricing` managed setting feeds your org's contracted per-model rates and discount multiplier into `/cost`, the status line, and telemetry instead of list price. Cost readouts finally line up with the invoice.
Sign in without an API key
`/login` → Anthropic Console now offers "Sign in with your Console account" alongside creating a key, so organizations that don't allow API keys can still authenticate and run Claude Code.
Remote MCP servers self-heal
In non-interactive (`-p`) and SDK sessions, remote MCP servers that dropped a connection used to stay dead until restart; they now reconnect automatically or report as failed. That's fewer silent hangs in your CI jobs and agent automation.
75 MB binary, plus a Linux crash fixed
The native binary is now zstd-compressed — about 75 MB instead of 340 MB on Linux x64 — shrinking install and auto-update downloads. Follow-up 2.1.245 fixes a startup crash on glibc 2.44 distros like Arch, CachyOS, and Fedora Rawhide.