Hermes · Preliminary Findings · Not decision-grade

Frontier Model Selection per Constellation

Date 2026-07-15 Author Hermes Claude facts live-verified (claude-api SSOT) Mapping inferred, not benchmarked GPT axis not researched

Methodology declared up front (rule #05 research-depth tiering): the model facts below are verified against the Anthropic model catalog (cached 2026-06-24, newer than training cutoff). The per-constellation mapping is reasoned from documented characteristics + known roles — not head-to-head benchmarked. This report deliberately stops short of a decision-grade arc.

§0Premise corrections

The originating question named "Fable 5 and Opus 5.6" and "GPT 5.6." Three corrections gate everything downstream:

There is no Opus 5.6. The Opus line tops out at Opus 4.8 (claude-opus-4-8). The current "5"-generation flagship is Fable 5 (claude-fable-5) — a distinct top tier, not an Opus point-release.
GPT is a different vendor and architecture. The whole Pantheon runs on Claude Code. Swapping a constellation to an OpenAI GPT model is a harness/architecture change, not a config swap — and current GPT specs can't be asserted from a stale (Jan 2026) prior without live research. Out of scope here.
Both Claude frontier tiers carry a 1M context window. "Smaller context window" and "Fable is context-limited" don't apply — Fable 5 and Opus 4.8 are both 1M.

§1Fable 5 vs Opus 4.8 — the real comparison

Dimension	Fable 5	Opus 4.8
Position	Most capable widely-released	Most capable Opus-tier
Context / max output	1M / 128K	1M / 128K
Price (in / out per MTok)	$10 / $50	$5 / $25 (half)
Thinking	Always on; can't disable	Adaptive; can disable
Turn latency	Minutes on hard tasks	Faster; fast mode ≈2.5× throughput
Data retention	30-day required (no ZDR)	Standard
Named strengths	Long-horizon autonomous runs; first-shot well-specified systems; enterprise deliverables (financial analysis, spreadsheets, docs); repo-history search; ambiguity; parallel sub-agent + async peer comms	SOTA agentic execution; knowledge work; memory; one-shot bug fixes; catches flakes; mid-session system messages; warmer writing
Security caveat	flag Cyber/bio classifiers false-positive on legit security tooling; bug-finding gains exclude security analysis	None

§2The key reframe

The intuition "Fable = planning, weak at engineering → switch to Opus for engineering" is inverted. Fable 5's largest gains are in long-horizon autonomous coding and first-shot system builds — it's the top tier at both planning and hard engineering. The real decision variable is not planning-vs-engineering. It is:

stakes × latency-tolerance × volume ÷ cost

Tier	Economics	Use for
Fable 5	2× price, minutes/turn	Rare, high-stakes, long-horizon, latency-tolerant work where a better outcome clearly justifies cost + wait
Opus 4.8	Half cost, fast mode	The fleet default workhorse — near-top capability, interactive-friendly
Sonnet 5	$3 / $15	High-volume lanes; near-Opus on coding/agentic
Haiku 4.5	$1 / $5, 200K ctx	Subagents, sweeps, simple tasks

§3Preliminary per-constellation mapping

Constellation	Primary role	Tier	Rationale
Athena	Governance, cross-fleet rulings	Fable 5	Rare, load-bearing verdicts; deepest reasoning wins, cost irrelevant at low volume
Proteus	Systems architecture, schema/DDL, roadmaps	Fable 5	"First-shot well-specified systems" is Fable's named sweet spot (Opus for routine DDL)
Plutus	Trading, contracts, financial analysis	Fable 5	"Enterprise deliverables — financial analysis, spreadsheets" is a named strength (Opus for ops)
Hephaistos	PR review / merge gate	Opus 4.8	Strong bug-finding + fast-mode throughput + cost; escalate hardest reviews to Fable
Hades	Credentials, vault, security scans	Opus 4.8 not Fable	Fable's cyber classifier false-positives on legit security work; bug-finding excludes security
Atlas	Infra/ops, skills & rules authoring	Opus 4.8	Careful critical-path authoring, high volume
Metis	CRM/scheduling architecture + build	Opus 4.8	Mixed design/build (Fable for large architecture arcs)
Hermes	Research + intelligence pipeline	Opus 4.8 + Fable + Sonnet	Opus controller; Fable for T-deep arcs; Sonnet subagents; Haiku sweeps. Volume drives cost efficiency
Apollo	Frontend / UI / visual	Opus 4.8 / Sonnet	Design instinct adequate; cheaper
Mnemosyne	Memory / knowledge graph	Opus 4.8	Pipeline/build work
pantheon-ops	n8n / hub automation	Sonnet 5 / Opus	Ops volume

Shape of it: Fable 5 for the 3–4 constellations whose output is rare, high-stakes, and long-horizon (Athena, Proteus, Plutus-analysis). Opus 4.8 as fleet default everywhere else. Sonnet 5 for high-volume/ops. Secondary lever — Opus 4.8 fast mode for interactive coordinator sessions, Fable 5 for overnight autonomous runs where latency is free.

Hard carve-out: keep Hades and all security-adjacent work off Fable 5 — cyber-classifier friction plus documented exclusion of security bug-finding.

§4Open before this becomes decision-grade

Cost envelope — no per-constellation volume/token estimates; the cost half of the formula is unquantified.
GPT / cross-vendor axis — not researched; would also weigh a Claude Code → alternate-harness architecture change.
Head-to-head benchmarks — mapping rests on vendor capability descriptions, not independent coding/agentic benchmark data.
Fable classifier friction — the Hades security exclusion is vendor-stated; real false-positive rate on fleet security tasks unverified.

Promote to a T-deep arc (parallel per-model sub-agent lanes, live-sourced benchmarks, cited verdict, cost model) only if a fleet-wide model-assignment ruling will hang on it.