Three versions have shipped since 2.1.305: 2.1.306 on October 21 at 02:28 UTC, 2.1.307 the same day at 20:47 UTC, and 2.1.308 on October 22 at 04:19 UTC. The middle one carries everything worth reading.
Two new things in 2.1.307
/claude-api api-version moves Claude API code to API version 2026-10-21. The timing is no accident: Anthropic released that version the day before, and it removes temperature, top_p and top_k outright. If you maintain integrations, the migration arrives as a command rather than as hand work.
There’s also a new OpenTelemetry event, claude_code.safety_event. It fires for every tool call that a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted on, or sent back. For teams running Claude Code across an organisation, that’s the number that was missing: not how much Claude did, but how often something intervened.
The fixes that matter in production
2.1.307 is mostly a long bug list, and several entries hit things people have been complaining about for weeks.
Requests were failing with a model-not-found error when CLAUDE_CODE_DISABLE_1M_CONTEXT was set and the model was an alias carrying the [1m] suffix, such as opus[1m]. MCP servers reported “needs authentication” under claude -p on macOS while the login keychain was locked; now they say why, and the next run after an unlock connects without a fresh sign-in. Also MCP: servers were reported as timed out, and stayed failed in headless runs, when the machine slept mid-connection.
HTTP MCP tool calls were being sent twice when the connection closed during the call — only read-only or idempotent tools get retried now. Plus a batch of “Autocompact is thrashing” errors in turns that read many PDF pages, images or other large tool results.
For cloud sessions: /autocompact fell back to the default window after a restart, and a conversation cleared with /clear sometimes came back once a cloud or self-hosted runner session restarted.
2.1.306 fixed two things: claude -p inside cloud sessions reporting “Not logged in” when started without the session’s environment variables, and an API error being swapped for “prompt is too long”, with a needless compaction, in sessions holding many MCP tools.
Three MCP entries in one bug list is a pattern
Five entries in 2.1.307 touch MCP, and three of those are connection failures that presented as something else: sleep as a timeout, a locked keychain as missing authentication, a dropped connection as a duplicate call. That’s the characteristic failure class of a layer moving from “works on my machine” to continuous operation. And claude_code.safety_event belongs to the same shift: anyone rolling out MCP servers and hooks across a company doesn’t need more features, they need numbers on what those features refused.