Make Fable 5 Run OpenAI’s GPT-5.6 Sol (So Neither Model Ships Its Own Mistakes)

TL;DR: Instead of arguing over whether Fable 5 or GPT-5.6 Sol is the better model, wire them together. Fable 5 plans and reviews, GPT-5.6 Sol builds, and the work loops between them until Fable approves it. It is one Claude Code skill, called on demand with /route, so it does not run on every prompt or burn your limits. Below is the from-scratch setup: the Codex plugin, pointing Codex at Sol, the skill file itself, and a demo prompt to try it on.

Everyone treats a new model launch as a leaderboard update: which one wins. That is the wrong question the day a lab drops something like GPT-5.6. The better question is what happens when you stop picking a side and make two rival labs’ models check each other’s work instead.

The setup: one boss, one worker, one loop

Fable 5 becomes the boss. It plans, and later it reviews, but it never writes the implementation itself. GPT-5.6 Sol is the worker: it gets the plan and builds it. When Sol hands the build back, Fable reads the diff like an adversarial reviewer looking for what Sol missed, sends specific fixes back, and the loop repeats until Fable actually approves it. Nothing ships until the boss signs off.

The whole thing lives in one Claude Code skill called /route, so it sits quiet until you actually need it.

Step 1: install the Codex plugin

This is what lets Claude Code hand work to Codex without leaving your session.

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

The last line is the one people skip. It is also the one that installs the Codex CLI itself if you do not already have it.

Step 2: point Codex at GPT-5.6 Sol

Update Codex so its model list is current, then confirm the exact model id, since ids shift as labs ship new releases.

codex update
codex /codex:status

Then create a .codex/config.toml at the root of your project:

model = "gpt-5.6-sol"

[profiles.reviewer]
model = "gpt-5.6-sol"
model_reasoning_effort = "xhigh"

Step 3: install the /route skill

This is a real Claude Code skill, not a loose markdown file. Save it as SKILL.md inside a route folder in your Claude Code skills directory, then reload plugins so it registers. The skill itself just is the loop: restate the task, write a plan, hand it to Sol with codex exec, review with /codex:review, send fixes back, repeat until approved.

Step 4: run it

/route "add input validation to the signup form and cover it with a test"

Fable writes the plan first. Then it dispatches the build to Sol. Then it reviews Sol’s work itself, the way an adversarial reviewer would, and if something is wrong, it goes back to Sol to fix before you ever see it.

The honest part

gpt-5.6-sol is the model id as shipped at the time this guide was written. If Codex rejects it, /codex:status will show you the current id to swap into the config file. The mechanism, plan then build then adversarial review, holds regardless of which exact model name Codex is pointed at.

The takeaway

Stop treating every model launch as a fight to pick a winner. The move that actually compounds is making the two best models in the industry check each other, one planning and reviewing, one building, looping until the work is clean. That is what /route wires up in about ten minutes.

I put the full setup, the config file, and the skill itself into a free guide. Comment ROUTE on the post and I will send it to you.