Production-grade reasoning that doesn't stall — --fallback-model

Run Opus at max effort, but auto-failover to Sonnet when Opus is overloaded. The flag that keeps cron jobs alive.

$ curl -s "https://hacker-news.firebaseio.com/v0/topstories.json" \
  | jq -r '.[0:30][]' \
  | xargs -P 10 -I {} curl -s "https://hacker-news.firebaseio.com/v0/item/{}.json" \
  | jq -s -r 'sort_by(-.score) | .[0:10] | .[] | "- [\(.score)] \(.title) — \(.url // "discussion")"' \
  | claude -p \
      --model opus \
      --effort high \
      --fallback-model sonnet \
      "Today's top HN stories. Identify the one quietly important but under-discussed. 5 sentences for an outsider."

1. [01] curlcurl … topstories.json …
   Same HN front-page fetch as recipe #1.
2. [02] claudeclaude -p
   Headless print mode. Required for --max-budget-usd, --fallback-model, --no-session-persistence.
3. [03] claude--model opus
   Pick the strongest model — opus, sonnet, haiku alias, or full ID like `claude-sonnet-4-6`.
4. [04] claude--effort high
   Reasoning effort: low / medium / high / xhigh / max. High = real extended thinking. Latency rises with effort.
5. [05] claude--fallback-model sonnet
   When opus returns 529 (overloaded), claude automatically retries with sonnet. Production-grade: the pipe never stalls. Only works with -p.

The under-discussed story is "The hidden cost of LLM batching" (score 388). While the M5 Pro benchmarks and the Rust vector DB get the upvotes, this one answers a question every team running production LLMs hits within their first month: how aggressive batching changes the latency-vs-cost tradeoff in a way that breaks SLA monitoring...

• --fallback-model only fires on overload (5xx), not on errors like rate limits.
• Test it by setting `--model haiku --fallback-model nonsense-model` — you should still get a haiku response, since haiku never overloads.