Enable enhanced model selection for agents

[don-petry]

Opus 4.8 has been released and is higher performing in a number of areas.

Research the optimal usage and price value of this new model and then enable it for those uses in the org agents. Enhance the agent harness to select the appropriate model and thinking level aligned with highest value (capability / cost) appropriate for the run.

https://www.anthropic.com/news/claude-opus-4-8

https://cursor.com/cursorbench

[don-petry]

Update — fold in Claude Fable 5 (new tier above Opus) + current engine state

Since this idea was filed for Opus 4.8, Anthropic has shipped Claude Fable 5 — a new tier above Opus. Both should be evaluated together here. Captured below so this discussion stays the canonical home for "research + enable new Claude models in our agents."

Models on the table (verified via the claude-api skill catalog)

Model	ID	Context	Max out	$/1M in	$/1M out	Notes
Claude Fable 5	claude-fable-5	1M	128K	$10	$50	Most capable; new tier above Opus. Same request surface as Opus 4.7/4.8 plus one breaking change: explicit thinking:{type:"disabled"} → 400 (omit the param instead). Min cacheable prefix 2048 tok.
Claude Opus 4.8	claude-opus-4-8	1M	128K	$5	$25	Most capable Opus; same API surface as 4.7 — a 4.7→4.8 move is a model-ID swap + prompt re-tune, no new breaking changes.
Claude Sonnet 4.6	claude-sonnet-4-6	1M	64K	$3	$15	Current deep/action tier.
Claude Haiku 4.5	claude-haiku-4-5	200K	64K	$1	$5	Current triage tier.

Adaptive thinking only on Fable 5 / Opus 4.8 (thinking:{type:"adaptive"}); budget_tokens and sampling params (temperature/top_p/top_k) are removed (400). Effort is controlled via output_config.effort (low|medium|high|xhigh|max).

Where the Claude engine sits today (scripts/engine.sh)

ENGINE_TRIAGE_MODEL = claude-haiku-4-5-20251001
ENGINE_DEEP_MODEL   = claude-sonnet-4-6
ENGINE_AUDIT_MODEL  = claude-opus-4-7      ← top tier still on 4.7
ENGINE_ACTION_MODEL = claude-sonnet-4-6
ENGINE_SINGLE_MODEL = claude-opus-4-7      ← single-reviewer still on 4.7

So the highest-capability slots are pinned to Opus 4.7 — neither Opus 4.8 nor Fable 5 is wired in anywhere. Per-tier fallback chains already exist (CLAUDE_*_MODEL_CHAIN), which gives us a clean place to stage a new model behind a fallback.

Dependency to flag

The tier routing is currently inert — see #195: every dev-lead intent routes through run_writer_with_fallback → ENGINE_ACTION_MODEL regardless of complexity, so the triage/deep/audit pipeline the ENGINE_LABEL advertises doesn't actually execute, and the rate-limit-based failover detection is broken. Enabling a higher-cost top tier (Fable 5 / Opus 4.8) only pays off once #195 is fixed and tasks are routed by capability. #195 is a prerequisite, not optional.

Related work

• Enhance dev-lead and pr-review engines for Gemini 3.5 models #382 — same idea for the Gemini 3.5 engine (parallel, non-Claude). Worth mirroring its structure.
• Initiative: Safe Release Strategy for Agentic Workflows (versioning · rings · canary) #495 — Safe Release Strategy initiative; any model bump to the self-hosting agents should ride the ring-staged rollout, not a blind flip.

Proposed next step

Convert this into a tracked implementation issue (capability/cost-aware selection across dev-lead + pr-review, Fable 5 + Opus 4.8 wired behind per-tier fallback, gated on #195). Filing that now and linking it back here.

[github-actions[bot]]

Weekly Update

What Changed

Two significant model releases since this discussion was created (May 31):

1. Claude Opus 4.8 (released May 28, 2026 — claude-opus-4-8):

   • Mid-conversation system messages — inject updated context without restarting, preserving prompt cache hits
   • Better long-horizon agentic coding — fewer compactions, better compaction recovery
   • Improved tool triggering — fewer skipped tool calls
   • Lower prompt cache minimum: 1,024 tokens (down from Opus 4.7's threshold)
   • Fast mode (research preview): 2.5x faster output at premium pricing
   • Adaptive thinking improvements: fewer wasted thinking tokens on simple turns

2. Claude Fable 5 (released June 9, 2026 — claude-fable-5):

   • First "Mythos-class" model — positioned above the Opus family
   • Pricing: $10/$50 per M tokens (input/output)
   • Can work autonomously for extended periods ("days of work")
   • Safety fallback: high-risk queries automatically route to Opus 4.8 (<5% of sessions)
   • Superior FrontierCode benchmark scores, even at medium effort level
   • More token-efficient than previous Claude versions on complex reasoning

Issue #553 now tracks the engine.sh integration. A new Ideas discussion (#562) proposes the specific implementation: Opus 4.8 as the new default Opus tier, Fable 5 as an optional "apex" tier for the audit/single-reviewer path on high-complexity PRs.

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	high	high	→
Impact	high	high	→
Urgency	med	high	↑ — both models now GA; issue #553 filed; competitive parity requires model currency

Recommendation

Advance — The models are released and available. Implement via #553 / Discussion #562. The engine.sh model chain architecture makes this a bounded, reversible change.

[don-petry]

Cross-link — related ideation output. The feature-ideation (BMAD Analyst) workflow now runs in this repo (adopted in #555) and independently surfaced the model-selection direction this thread was waiting on:

• 💡 Fable 5 Apex Tier + Opus 4.8 Upgrade for engine.sh Model Chains #562 — Fable 5 Apex Tier + Opus 4.8 Upgrade for engine.sh Model Chains
• 💡 Complexity-Aware Dynamic Model Routing for Cost-Optimized PR Review #564 — Complexity-Aware Dynamic Model Routing for Cost-Optimized PR Review

This discussion (#413) remains the canonical home for "research + enable new Claude models in our agents", and #553 is the tracked implementation issue (gated on #195).

[don-petry]

Implemented — this idea has shipped. The capability/cost-aware model enablement requested here was tracked as #553 and implemented via #569 (merged 2026-06-11). scripts/engine.sh now runs the new tiers: ENGINE_DEEP_MODEL=claude-opus-4-8, ENGINE_AUDIT_MODEL=claude-fable-5, ENGINE_SINGLE_MODEL=claude-fable-5, with per-tier capability-aware fallback chains (CLAUDE_*_MODEL_CHAIN). Opus 4.8 and Fable 5 are wired in across the dev-lead + pr-review engines. Safe to close this discussion as resolved. (Follow-on idea #564 — complexity-aware dynamic routing — remains a separate, open initiative.)