💡 Opus 4.8 Engine Upgrade: Dynamic Workflows + 4x Honesty for Review Quality

[github-actions[bot]]

Summary

Upgrade engine.sh model references from Opus 4.7 to Opus 4.8 (shipped May 28, 2026) to leverage two key capabilities: Dynamic Workflows (model-internal parallel subagent orchestration for large diffs) and a 4x reduction in missed flawed code. Fast Mode pricing dropped from $30/$150 to $10/$50 per MTok, making speed-optimized Opus viable for latency-sensitive review tiers.

Market Signal

Opus 4.8 benchmarks (May 28, 2026 release):

• SWE-bench Pro: 69.2% (up from 64.3% on Opus 4.7)
• Terminal-Bench 2.1: 74.6% (up from 66.1%)
• GPQA Diamond: 93.6%
• Honesty: 4x less likely to miss flawed code; 17x less likely to produce dishonest summaries of agentic work
• Dynamic Workflows: Model-internal ability to plan, spawn parallel subagents, and synthesize results — Opus 4.8's signature capability
• Fast Mode: $10/$50 per MTok (down from $30/$150 on Opus 4.7) — a 3x cost reduction that fundamentally changes the economics of speed-optimized review

User Signal

• Existing discussion 💡 Fable 5 Apex Tier + Opus 4.8 Upgrade for engine.sh Model Chains #562 proposed Opus 4.8 integration but predated the actual release
• The engine.sh tier architecture (rubber-duck → triage → deep) is already designed for model-per-tier routing
• model-pricing.tsv provides the cost data infrastructure for pricing updates
• The skill eval framework ([Phase 1] Build the deterministic eval scorer + offline tests (triage) #583, [Phase 2] Extend the harness to prompts/deep-review.md via an LLM-judge scorer #585, [Phase 2] Strict-improvement gate + manual propose->validate->PR runbook with regression-rejection proof #586) was recently shipped, providing the regression-testing infrastructure needed for safe model upgrades

Technical Opportunity

engine.sh already supports model selection per tier. Updating model references and adding Opus 4.8 pricing to model-pricing.tsv is a low-risk change. The Dynamic Workflows capability requires no code changes — it activates automatically when Opus 4.8 encounters complex multi-part tasks, improving deep-review quality without prompt engineering changes. Fast Mode viability at $10/$50 opens a new "express" tier option for latency-sensitive reviews (small hotfix PRs, security patches).

Assessment

Dimension	Score	Rationale
Feasibility	high	Model reference update + pricing data; eval framework provides safety net
Impact	high	4x honesty improvement directly improves review quality; Fast Mode enables new tier
Urgency	high	Opus 4.8 is shipped and available; delayed upgrade means missing quality improvements

Adversarial Review

Strongest objection: Model upgrades risk regression — prompts were tuned for Opus 4.7 behavior. Dynamic Workflows is model-internal and not controllable, meaning review output structure could change unpredictably.

Rebuttal: The skill eval framework (#583, #585, #586) exists precisely to catch regressions from model upgrades. Run the triage and deep-review eval suites against Opus 4.8 before promoting. Dynamic Workflows is triggered by task complexity — simple reviews won't activate it. The 4x honesty improvement alone justifies the upgrade even without Dynamic Workflows.

Suggested Next Step

Run the triage and deep-review skill eval suites against Opus 4.8, compare quality scores to Opus 4.7 baseline, and update model-pricing.tsv with Opus 4.8 rates including the new Fast Mode pricing.

[don-petry]

I agree with the approach, but do not want to enable the fast tier as we do not have any need for an Express PR review capability. The bottleneck in PR reviews is not the speed of the model.

[github-actions[bot]]

📋 Initiative planned by the BMAD Scrum Master (Bob).

Epic #835 — Opus 4.8 engine upgrade: validate the deep-review tier, finish the reference rollout, and keep cost data accurate

3 stories created (inert — labelled initiative, NOT initiative:auto):

• [Phase 1] Validate Opus 4.8 >= Opus 4.7 on the deep-review held-out eval suite (regression gate) #836 (M) — [Phase 1] Validate Opus 4.8 >= Opus 4.7 on the deep-review held-out eval suite (regression gate)
• [Phase 2] Finish the Opus 4.8 reference rollout in engine.sh, aw.sh, and the dev-lead spec #837 (S) — [Phase 2] Finish the Opus 4.8 reference rollout in engine.sh, aw.sh, and the dev-lead spec
• [Phase 1] Confirm Opus 4.8 standard pricing in model-pricing.tsv and lock it with a regression test #838 (S) — [Phase 1] Confirm Opus 4.8 standard pricing in model-pricing.tsv and lock it with a regression test

Open questions for review:

• Advisory: docs/dev-lead/spec.md's model table is stale beyond the Opus 4.7 references (e.g. it lists ENGINE_DEEP_MODEL = claude-sonnet-4-6 while engine.sh sets claude-opus-4-8, and Gemini/Copilot columns may also drift). Story 2 reconciles the Opus references and the rows it must touch to stay consistent — confirm whether a fuller spec.md model-table audit should be folded in here or tracked as a separate doc-hygiene task.

---

Review the epic and its sub-issue DAG, adjust as needed, then add initiative:auto to epic #835 to hand it to initiative-driver for auto-implementation.