"When the Default Model Is Rejected, Fall Back Once to the Previous One" — Claude Code v2.1.286 Closes the "Every-Turn Failure" in Unattended Runs

On September 30, 2026, Anthropic released v2.1.286 of its coding agent, Claude Code. Rather than flashy new features, the focus is on quiet fixes that eliminate the kind of accidents where an agent running unattended "silently stops." The most impactful of these is a single change to how Claude Code behaves when the default model is rejected by the API.

Drop a rejected model to the "previous model" in the same tier and keep going

Until now, if the Anthropic API refused the specified default model (or the model that an alias resolved to), that session would fail on every turn. A model name that merely failed to go through temporarily was enough to bring the work itself to a halt.

v2.1.286 changes this behavior. The official changelog puts it this way: "Fixed an issue where Claude Code would fail every turn when the Anthropic API rejected the default or alias-resolved model. Claude Code now retries once with the previous model in the same tier."

If a human is watching, they can see the error and pick a different model. But when you run unattended via cron or CI, the presence or absence of this "fall back once to the previous model" behavior is what separates a job that finishes from one that wipes out entirely. It's a fix clearly aimed at automated operation.

Pile-ups of approval requests now show a "N of M" counter

Another change is how approval prompts appear. When multiple permission requests overlap, the prompt now shows a count like "2 of 5."

In recent versions, Claude Code has shifted the default for interactive sessions toward auto mode. As decisions proceed automatically, it becomes harder to tell "how many approvals are currently queued up." The counter is a small peephole that lets a human grasp that backlog. The "N more" rows inside the full-screen list are now operable with the mouse as well, so you can jump straight to the far end of a long list.

When an organization turns off remote control, the session is cut too

There's a governance fix as well. Previously, even if an organization disabled Remote Control via policy, sessions that were already established could remain connected. In v2.1.286, the moment the policy is turned off, the session is notified and disconnected.

It plugs the situation where "the admin shuts the tap, but water keeps leaking from an outlet that was already open." When an enterprise hands out unattended agents, it should be able to assume that revoking a permission actually drops the live connection at that instant.

Login glitches cleaned up too

Several smaller authentication bugs were also addressed. The main ones are as follows.

• Fixed an issue where, when GCP/AWS credentials (gcpAuthRefresh / awsAuthRefresh) expired, multiple processes and IDE extensions would each open their own login browser.
• Fixed an issue on macOS where a session kept showing "Not logged in" or "Login expired" even after a successful /login in another window.

The VS Code extension got attention as well, adding "bookmarks" to save Claude's responses, option previews for question cards, and extra context in messages that indicate terminal output or browser use.

What this move means on the ground

What runs through v2.1.286 is a push toward "not breaking silently, even with no human present." Not stalling on a model rejection, making the approval backlog visible, reflecting permission revocation in the connection — all of these are adjustments for the people keeping agents running. The checkpoints for adopters are narrow: automated jobs that use a default model or alias should verify, on the assumption that the fallback drops to "the previous model in the same tier," that running an unexpected model won't cause problems.

At the same time, a fallback is double-edged. Because processing continues even when the model switches, the anomaly of "the model you were supposed to use isn't going through" can hide behind a normal, successful completion rather than surfacing as an error. One could reasonably argue that you should make such quiet switchovers catchable in your monitoring logs. It's worth keeping in mind that, behind the convenience, anomaly detection gets one notch harder.

It isn't flashy, but it's an update that firms up the foundation for unattended operation. The more a pipeline is built on auto mode and fallbacks, the more directly both the benefits and the caveats come into play.

References: Claude Code Official Changelog / Releasebot: Claude Code Updates / Havoptic: Claude Code Release Summaries