Claude Code Errors on Sonnet 5.5: How to Keep Working with --fallback-model
Back to the blog
Articleclaudetroubleshootingmodelos de IAagente de programaçãodev tools

Claude Code Errors on Sonnet 5.5: How to Keep Working with --fallback-model

MafraOctober 27, 20264 min read

You send a request in Claude Code and get a failure instead of an answer. On Oct 27, 2026, Anthropic's status page logged the incident "Elevated errors for Claude Sonnet 5.5": elevated errors on requests to Sonnet 5.5.

Short answer: check status.claude.com, and if the problem is on their side, run claude --fallback-model sonnet,haiku or switch models with /model. Today's incident is already resolved, but the steps apply to the next one.

Flowchart: confirm the incident on status.claude.com, turn on --fallback-model, switch models with /model and, to make it permanent, save fallbackModel in settings
Three steps to keep working, and the branch to make the chain permanent.

What happened to Sonnet 5.5 on Oct 27, 2026?

Anthropic's status page posted three updates the same day, all in UTC:

Time (UTC)StatePublished text
09:14Investigating"We are investigating elevated errors on requests to Claude Sonnet 5.5."
09:38Monitoring"A fix has been deployed for the issue affecting Claude services and we are monitoring for recovery."
09:56Resolved"The issue affecting Claude services has been resolved."

That is about 40 minutes from investigation to resolution. The source is the status.claude.com/history.atom feed, checked on Oct 27, 2026.

How do you tell if the error is yours or Anthropic's?

Open status.claude.com before touching any configuration. If an incident is open for the model you use, the error is on their side, and reinstalling, logging in again or clearing caches fixes nothing.

If the page is clean, the problem is something else: authentication, billing, usage limits or network. Those cases do not trigger the fallback, as the table below shows.

How do you make Claude Code switch models on its own?

Use the --fallback-model flag with a comma-separated list of models. According to the official CLI reference, it enables automatic fallback when the primary model is overloaded or not available, trying the models in the order you listed.

claude --fallback-model sonnet,haiku

Per the model configuration docs, this is how it behaves:

  • The chain holds at most 3 models after removing duplicates. Extra entries are ignored.
  • Each entry accepts a model name or an alias. "default" expands to the default model.
  • The switch lasts for the current turn only. On your next message, Claude Code tries the primary model again.
  • When it switches, it shows a notice. /status does not display the chain, so that notice is the first sign the fallback is active.
  • Subagents use the chain too. Your session's model does not change.

When does the fallback trigger, and when does it not?

SituationSwitches model?
Model overloadedYes
Model unavailableYes
Other non-retryable server errorYes
Authentication errorNo
Billing errorNo
Usage limit (rate limit)No
Request too large or network errorNo

That is why the fallback is not a cure for "I hit my limit". It covers provider failure, not missing quota.

How do you switch models mid-session?

Type /model sonnet (or another alias) to switch right away without leaving the conversation. With no argument, /model opens the picker. One detail that changes the outcome, according to the docs:

  • Enter switches the model and saves it as your default.
  • s switches for this session only and leaves your default alone.

Typing /model <name> directly behaves like Enter, so it writes to your settings. If the switch is only to get through the incident, open the picker and press s, or you will still be on the backup model tomorrow without noticing.

How do you keep the fallback chain always on?

Save the fallbackModel field in your settings, as a list:

{
  "fallbackModel": ["sonnet", "haiku"]
}

It applies to every new session. If you pass --fallback-model on the command line, the flag takes precedence over the setting.

Heads up: the backup model is usually smaller. The docs say that during context compaction, Claude Code will not fall back to a model with a smaller context window than the primary one. If every fallback is smaller, you see the original error and can retry.

What if you do not want to depend on a single provider?

The fallback helps, but it stays inside the same provider. Verboo Code lets you pick from more than one model, with unlimited tokens, and runs in the terminal like Claude Code. Worth having as a second option for days like this one.

Enjoyed this article?
Share knowledge with your network.
// Read also

Related articles