Switch Opus 5.5 and Sonnet 5.5 mid-session: Sume jobs keep running

You can switch between Opus 5.5 and Sonnet 5.5 in the middle of a Claude Code session that is waiting on Sume jobs, because a Sume job belongs to your workspace on Sume's side, not to the model that submitted it. Claude Code 2.1.287 also fixed a bug where switching with /model or opusplan rewrote earlier MCP tool announcements, which could drop earlier extended thinking.

The fix is in the Claude Code changelog entry for October 1, 2026; the model ids and prices are from Anthropic's models overview. Both were read 2026-10-01.

What exactly did 2.1.287 fix?

The changelog line: switching between Opus 5.5 and Sonnet 5.5 with /model or opusplan rewrote earlier MCP tool announcements, which could drop earlier extended thinking. So before the update a switch could change how earlier MCP tools were presented to the model. After it, the earlier announcements are left alone.

Which two models are involved?

Anthropic's overview lists both with a 1M-token context window and 128K max output. The Claude Code changelog adds that Sonnet 5.5 became the default Sonnet on September 28.

Where does a Sume job live while I switch?

Sume's jobs page says generation endpoints create durable jobs and to store the job id so work can be recovered after process restarts. A job in queued or processing keeps going whichever model is in the chat, and it keeps billing. A model switch changes who reads the result, not whether the job runs.

How does the new model pick up an in-flight job?

Paste or restate the job ids, or let the new model call jobs_list, which is a read tool. Then jobs_wait with those ids; on remote MCP a wait slice is capped at 55 seconds, and the docs say to retry the wait with the same ids rather than resubmit the paid create. jobs_result returns the outputs once a job is completed. The tools and gates page lists all three.

Should I switch before or after a paid submit?

Switching between submit and wait is safe for the job. What the switch cannot do is un-send a call, so if you want the cheaper model to do the polling, submit with the stronger one, note the ids, then switch. Keep idempotency_key stable for any retry.