Quick Summary
- Claude Fable 5 carries judgment, voice, and ambiguity.
- GPT-5.6 Sol through Codex carries execution volume.
- The source corpus, acceptance criteria, and human review gate matter more than a model leaderboard.
- The Current Split
- What Fable Carries
- What Sol Carries
- Claude Code and Codex CLI
- The Source-Grounded Workflow
- How to Choose
The Current Split
The current production stack uses Claude Fable 5 and GPT-5.6 Sol through Codex as complementary workers.
Fable carries work where a wrong tone, weak distinction, or missed implication changes the result. Sol carries bounded execution where the contract, files, and acceptance tests make success observable.
This is an operating split, not a universal benchmark. Models change. The durable assets are the source material, written standards, retrieval layer, correction history, and review gate.
What Fable Carries
Fable is the judgment and voice lane.
Use it for:
- positioning and editorial judgment
- first-person voice and client-facing nuance
- resolving ambiguity in a brief
- evaluating whether an argument says something worth publishing
- final taste review when several technically valid drafts exist
Fable still starts from sources. Judgment does not excuse unsupported claims.
What Sol Carries
Sol is the execution-volume lane.
Use GPT-5.6 Sol through Codex for:
- repository and vault search
- research extraction from provided sources
- formatting and structured transformations
- link and claim sweeps
- code changes with explicit tests
- repeatable article builds
- evidence-backed audits and receipts
Sol works best when the task names the inputs, target files, constraints, and proof of completion.
Claude Code and Codex CLI
The model and the execution tool are separate choices.
Claude Code and Codex CLI both operate against local files and tools. The current stack uses both, with Codex CLI named as the primary execution surface and Claude Code retained for Fable-driven judgment and voice work.
The shared vault is the memory plane. Neither tool's private session history is the only record.
The Source-Grounded Workflow
- Define the claim, audience, and acceptance criteria.
- Gather the current source text.
- Let Sol extract, compare, structure, and execute the bounded work.
- Let Fable judge positioning, voice, and unresolved ambiguity.
- Verify links, numbers, quotes, and runtime behavior deterministically.
- Persist the accepted output and correction record outside the vendor session.
The workflow can use one model for several steps when the task is small. The roles matter more than ceremony.
How to Choose
Choose Fable when the work turns on judgment.
Choose Sol when the work turns on execution against a clear contract.
Use both when a deliverable needs volume and taste. Keep the person at the gate for consequential claims and external side effects.
Do not choose from remembered pricing tables or old capability rankings. Check the current provider plan when cost affects the decision, then measure the full workflow: source preparation, model time, human editing, failures, and rework.
The practical rule is short: Fable carries judgment and voice. Sol carries execution volume.