◉CLAUDE LABJP
●WATERMARK — Since 11/16, text from Opus 4.6, Sonnet 4.5 and other models carries a watermark; non-default temperature and similar settings may drop it●CC 2.1.325 — Claude Code is still on 2.1.325 (11/13), with no newer release yet●SONNET 4.5 — Retires on the Claude API on 11/30, 13 days left. Sweep model IDs in code and saved settings in one pass●AUTOMATION — A user reports recurring AI jobs set up without run counts or a spending cap, burning through a weekly allowance●CONNECTOR — If connectors vanish only in new sessions, compare the /mcp list against an older session first●NEW — Cowork Feels Worse? Run the Same Request in Three Fresh Sessions●WATERMARK — Since 11/16, text from Opus 4.6, Sonnet 4.5 and other models carries a watermark; non-default temperature and similar settings may drop it●CC 2.1.325 — Claude Code is still on 2.1.325 (11/13), with no newer release yet●SONNET 4.5 — Retires on the Claude API on 11/30, 13 days left. Sweep model IDs in code and saved settings in one pass●AUTOMATION — A user reports recurring AI jobs set up without run counts or a spending cap, burning through a weekly allowance●CONNECTOR — If connectors vanish only in new sessions, compare the /mcp list against an older session first●NEW — Cowork Feels Worse? Run the Same Request in Three Fresh Sessions
Articles/Claude Code
⟐ Claude Code/2026-06-12Intermediate

A Three-Tier fallbackModel Setup for Claude Code — Keeping Unattended Runs Alive Through Overload Mornings

How I run Claude Code with a three-tier fallbackModel chain so overnight batches survive overload errors: logging which model actually ran, measuring quality drift on fallback days, and pairing it with deny rules.

claude-code134fallbackmodelautomation110scheduled-jobsreliability18

✦ Premium Article

One morning this June, my Crashlytics triage report arrived empty. As an indie developer I rely on a headless Claude Code run (claude -p) at 6 a.m. to classify the previous day's crash logs for my apps. That day, the log showed three overloaded_error responses (HTTP 529) in a row — the retry limit had been exhausted and the job had simply given up. It was a blunt reminder of how fragile a retry-only design really is. Unaddressed crashes feed directly into lower AdMob revenue and worse store reviews, which makes this triage the one batch in my operation I least want to lose.

When I tallied the previous 30 days of logs, the morning batch had come up empty on 4 of them — roughly a 13% loss rate. When you are sitting at the keyboard, "wait a bit and try again" solves this. Unattended, no amount of waiting wins against an overload that lasts longer than your retry window. So I moved to a three-tier setup using the fallbackModel setting that recently landed in Claude Code, switching models instead of merely retrying one. This article is a record of that design and what I learned actually operating it.

Where a retry loop alone stopped working

My pre-migration script was the usual exponential backoff:

#!/bin/bash
# Before: retries against the same model only. A long overload kills every attempt.
PROMPT_FILE="$HOME/ops/crashlytics_triage_prompt.md"
for i in 1 2 3; do
  if claude -p "$(cat "$PROMPT_FILE")" > /tmp/triage_result.md 2>/tmp/triage_err.log; then
    exit 0
  fi
  sleep $(( i * 60 ))  # 1 min -> 2 min -> 3 min
done
echo "triage failed after 3 attempts" >&2
exit 1

The flaw is that every retry queues up against the same congestion on the same model. A 529 is server-side overload; some mornings it clears in minutes, other mornings it lingers for half an hour or more. Looking back at the four failed days, all three retries had been spent within six minutes of the first failure — well inside the congestion window, every single time.

What changes with a fallbackModel array

Claude Code's fallbackModel accepts an array of up to three models in settings.json. When the primary cannot respond — overload, model unavailable — execution switches to the next model in line and keeps going. If retrying is rejoining the same queue, falling back is walking over to a different counter that happens to be open. For unattended runs, that difference is decisive.

// After: .claude/settings.json — switch models under overload instead of re-queueing
{
  "model": "claude-fable-5",
  "fallbackModel": ["claude-opus-4-8", "claude-sonnet-4-6"]
}

That is the entire change — the calling script stays untouched. In the 30 days since migrating, the fallback chain fired on 3 mornings, and the number of empty-report days dropped to zero. You cannot prevent 529s, but you can prevent them from costing you results.

One caveat from real operation: fallbackModel responds to overload and availability errors, not to timeouts. With an always-on adaptive-thinking model as the primary, even simple-looking tasks can think longer than expected, which makes wall-clock time harder to predict. I widened my batch-level timeout to 1.5x its previous value. Fallback and timeout are separate insurance policies and need to be designed separately.

✦

Thank you for reading this far.

Continue Reading

What follows includes implementation code, benchmarks, and practical content we hope you'll find useful. This site runs without ads — server and development costs are supported entirely by members like you. If it's been helpful, we'd be truly grateful for your support.

WHAT YOU'LL LEARN
✦A working Fable 5 → Opus 4.8 → Sonnet 4.6 fallbackModel chain and the criteria behind each slot
✦A bash implementation that extracts the executing model from the stream-json init message and logs it to CSV
✦A measured 90% classification-agreement rate on a Sonnet 4.6 fallback day, and the two operational rules it produced
Secure payment via Stripe · Cancel anytime
✦

Unlock This Article

Get full access to the rest of this article. Buy once, read anytime. This site is ad-free — your support goes directly toward keeping it running.

or
Unlock all articles with Membership →
Share

Thank You for Reading

Claude Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • ✦Copy-paste ready implementation code
  • ✦New advanced guides published daily
  • ✦$5/mo or $10 for lifetime access
View Membership →

Related Articles

⟐ Claude Code2026-04-30
Composing Claude Code Skills — Building Automation a Single Skill Can't Reach
A practical guide to going beyond one-skill-at-a-time and composing multiple Claude Code Skills into real workflows. Pipeline, branching, and meta-skill patterns with full implementation examples and the failure modes you should design against.
⟐ Claude Code2026-03-28
Running Claude Code × Playwright E2E Tests in Production — Pitfalls and Decision Criteria
Implementation patterns for running E2E tests with Claude Code and Playwright in production, plus how to measure whether the investment pays back. Includes the Cloudflare Workers and Pages preview deployment pitfalls, and a decision framework that indie developers can use to judge whether E2E tests are worth the investment.
⟐ Claude Code2026-03-24
How to Automate Game Development with Claude Code × unity-mcp — A Complete Workflow from Concept to Release
Learn how to combine Claude Code with unity-mcp to automate Unity game development. The hcg-workflows skill set provides an 8-phase workflow from planning to deployment.