💡 Fable 5 Apex Tier + Opus 4.8 Upgrade for engine.sh Model Chains

[github-actions[bot]]

Summary

Add Claude Opus 4.8 (claude-opus-4-8) and Claude Fable 5 (claude-fable-5) to engine.sh model configuration. Opus 4.8 replaces Opus 4.7 as the audit/single-reviewer model with improved long-horizon agentic coding and tool triggering. Fable 5 becomes an optional "apex" tier for the most complex reviews, activated only when PR complexity exceeds a threshold.

Market Signal

Anthropic released Opus 4.8 on May 28, 2026 with: mid-conversation system messages (preserves cache hits in long agent sessions), adaptive thinking improvements (fewer wasted thinking tokens), better tool triggering, and a lower cache minimum (1,024 tokens). Fable 5 launched June 9 as the first Mythos-class model — $10/$50 per M tokens, capable of multi-day autonomous tasks, with automatic safety fallback to Opus 4.8 on high-risk queries (<5% of sessions). Both are available now on the Claude API.

User Signal

Issue #553 explicitly requests enabling Fable 5 and Opus 4.8 with capability-aware selection. Issue #382 requested Gemini 3.5 support, showing ongoing demand for model currency. The engine.sh architecture already supports per-tier model chains (CLAUDE_TRIAGE_MODEL_CHAIN, CLAUDE_DEEP_MODEL_CHAIN, etc.) and cross-provider fallback.

Technical Opportunity

engine.sh already defines per-tier model chains and supports claude/gemini/copilot providers. Adding new model IDs requires updating set_engine_config() — the architecture was designed for exactly this kind of extension. Key Opus 4.8 features for the review pipeline:

• Mid-conversation system messages — inject updated CI status without restarting the full prompt, preserving cache hits
• Better compaction handling — long agentic traces stay on task with fewer derailments
• Lower cache minimum (1,024 tokens) — more prompts qualify for caching
• Fast mode — 2.5x faster output at premium pricing for time-sensitive reviews

Assessment

Dimension	Score	Rationale
Feasibility	high	engine.sh model chain architecture was designed for model upgrades; change is ~20 lines in set_engine_config() plus pricing rows
Impact	high	Better review quality from Opus 4.8's improved tool triggering + Fable 5's superior complex reasoning; mid-conversation system messages reduce cost
Urgency	high	Both models released in the last 2 weeks; issue #553 already filed; competitive parity requires model currency

Adversarial Review

Strongest objection: Fable 5 at $10/$50 per M tokens is ~2x Opus 4.7 pricing. Unrestricted use could balloon the token budget. The safety fallback to Opus 4.8 on sensitive topics could cause inconsistent behavior in security-focused reviews.

Rebuttal: Cost control comes from the existing tier architecture — Fable 5 would only activate for the audit tier on high-complexity PRs, not for triage or routine deep review. The safety fallback is actually beneficial for security reviews, routing sensitive code analysis to the more conservative Opus 4.8. The per-tier model chain pattern means adding Fable 5 to CLAUDE_AUDIT_MODEL_CHAIN is a minimal, reversible change. Update model-pricing.tsv with effective-dated rows to maintain cost reporting accuracy.

Suggested Next Step

Update set_engine_config() in scripts/engine.sh to use claude-opus-4-8 for all current claude-opus-4-7 slots, add claude-fable-5 as first entry in CLAUDE_AUDIT_MODEL_CHAIN and CLAUDE_SINGLE_MODEL_CHAIN, and update scripts/lib/model-pricing.tsv with new pricing rows.

[don-petry]

Cross-link. Canonical discussion for new-Claude-model enablement is #413 ("Enable enhanced model selection for agents"); tracked implementation is issue #553 (gated on the inert tier-routing bug #195). Fold this proposal's direction into that work rather than tracking it standalone.

[github-actions[bot]]

Weekly Update

What Changed

Implementation complete — PR #569 merged the Fable 5 and Opus 4.8 upgrade to engine.sh model chains. The core proposal is now implemented.

New cost signals require attention:

1. Fable 5 free inclusion ends June 22, 2026. After that date, Fable 5 counts as 2x usage on Claude subscription plans, consuming Agent SDK credits at double the Opus 4.8 rate. The audit tier chain (CLAUDE_AUDIT_MODEL_CHAIN=claude-fable-5,claude-opus-4-8,claude-opus-4-7) should have cost-aware gating that downgrades to Opus 4.8 when the budget is tight.
2. June 15 billing split (2 days away): Agent SDK usage moves to a separate monthly credit pool. Fable 5's $10/$50/MTok pricing makes budget awareness 2x more critical. See Discussion 💡 Agent SDK Credit Budget Circuit Breaker for Billing-Aware Orchestration #635 for the circuit breaker proposal.
3. SWE-bench data (June 2026): Fable 5 scores 95.0% on SWE-bench Verified and 80.3% on SWE-bench Pro. This is a 6.4-point improvement over Opus 4.8 (88.6%) but at 2x the cost — the cost-per-point ratio confirms that Fable 5 should be reserved for truly complex reviews, exactly as implemented.

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	high	high (implemented)	→
Impact	high	high	→
Urgency	high	med (implemented, but cost gating needed)	↓

Recommendation

Keep watching. Core implementation is done. Next priority: ensure the Fable 5 audit tier interacts properly with the budget circuit breaker (#635) and batch API (#639) proposals. After June 22, monitor Token Cost Observatory data for Fable 5 cost impact and adjust the model chain order if Fable 5 costs exceed the quality benefit.