Frontier Model Routing Strategy

July 27, 2026 · View on GitHub

Status: accepted and implemented Decision date: 2026-07-27

Decision

Claude Octopus uses Opus 5 as its premium lead model, GPT-5.6 Sol as the independent coding/review peer, and Sonnet 5 as the standard Claude seat. Fable 5 remains an explicit capability escalation rather than an automatic default.

Fresh configurations adopt the new roster. Existing environment pins, session overrides, and providers.json settings retain precedence and are not silently rewritten.

Roster

ModelDefault jobStandard price per MTok (input/output)
Claude Opus 5architecture, planning, security reasoning, final judgment$5 / $25
GPT-5.6 Solimplementation, terminal work, independent code review$5 / $30
GPT-5.6 Terrabalanced Codex alternative$2.50 / $15
GPT-5.6 Lunabudget Codex alternative$1 / $6
Claude Sonnet 5standard Claude orchestration and synthesis$3 / $15
Claude Haiku 4.5budget Claude work$1 / $5
Claude Fable 5opt-in judgment-class escalation$10 / $50

Opus 5 and Fable 5 are both Anthropic-family models. Agreement between them does not count as independent provider diversity.

Routing rules

  1. Start generic, mergeable work with one capable owner.
  2. Add another model only for a distinct job: research, implementation, adversarial review, security review, or final verification.
  3. Explicit debate, council, squeeze, and multi-model commands still fan out because disagreement is their intended output.
  4. Model choice never changes permissions, repository rules, or quality gates.
  5. User and project configuration always beats release defaults.

Role defaults:

  • architect, strategist, security-reviewer, implementer-heavy: current Opus, preferring Opus 5 on Claude Code v2.1.219+.
  • implementer, code-reviewer: GPT-5.6 Sol.
  • synthesizer: current Sonnet, preferring Sonnet 5 on Claude Code v2.1.197+.
  • researcher: Antigravity, retaining an independent research role.

Fallbacks

  • Opus: Opus 5 → Opus 4.8 → Opus 4.7 → Opus 4.6.
  • Sonnet: Sonnet 5 → Sonnet 4.6.
  • Fable refusal/security fallback: Opus 5, overridable with OCTOPUS_FABLE5_FALLBACK_MODEL.
  • GPT-5.6 requires Codex CLI v0.144.0 or newer.

Prompt policy

For Opus 5 and Sonnet 5, prompts should state the goal, relevant context, boundaries, reasons for unusual constraints, and checkable acceptance criteria. Avoid duplicated reminders, all-caps emphasis without a real compliance need, token countdowns, and requests to reveal hidden reasoning. The runtime policy block is skills/blocks/frontier-model-routing.md; Fable-specific constraints remain in skills/blocks/fable5-prompting.md.

Sources