Feature request: configurable safety-fallback target — Fable 5 should fall back to Opus 5, not Opus 4.8

Feature request

Make the safety-classifier fallback target configurable on the first-party API, or chain it through the newest Opus: Fable 5 → Opus 5 → Opus 4.8, instead of sending cybersecurity-flagged requests straight to Opus 4.8.

Current behavior (v2.1.219+)

Per the automatic model fallback docs, when a safety classifier flags a request on Fable 5:

• Biology-flagged requests re-run on Opus 5 ✅
• Cybersecurity-flagged requests re-run on Opus 4.8, and this target is not configurable on the Anthropic API / subscription auth

ANTHROPIC_DEFAULT_OPUS_MODEL already redirects the fallback target — but only on Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry ("flagged requests re-run on that model for every category that has a fallback"). On the first-party API it only changes what the opus alias resolves to.

Why this matters

• Fable 5 carries additional dual-use safety measures, so its cybersecurity classifiers flag more aggressively than Opus 5's. A meaningful share of Fable-flagged requests (benign security-adjacent work: defensive tooling, audits, networking code) would be served fine by Opus 5 — but the current routing skips it entirely.
• Users who deliberately run the newest models end up silently downgraded to a previous-generation Opus, and because the session persists on the fallback model, --resume/--continue sessions stay on Opus 4.8 indefinitely unless the user notices and runs /model.
• Workloads in offensive/defensive security trip this frequently — sometimes on the first request of a session (CLAUDE.md context alone can trigger it), so the fallback target is effectively a second default model for these users.

Requested change (any of these would resolve it)

1. Chain the cybersecurity fallback: Fable 5 → Opus 5, and only continue to Opus 4.8 if Opus 5 also flags the request.
2. Honor ANTHROPIC_DEFAULT_OPUS_MODEL for fallback-target resolution on the first-party API, matching the documented Bedrock/Agent Platform/Foundry behavior.
3. A settings key for the refusal-fallback target (validated to the Opus family, as the third-party path already does).

Secondary ask: an opt-in "return to the original model after the flagged turn" setting, so a fallback is a one-turn detour (like fallbackModel chains already are for availability fallback) instead of a persistent session switch.

Environment

• Claude Code v2.1.219, macOS (darwin 25.6.0)
• Subscription auth (no API key), model claude-fable-5[1m]
• switchModelsOnFlag at default true