Don't Marry the Model

Everyone is switching models again!

Fable got a stay of execution until Sunday! ChatGPT just dropped Sol, Terra and Luna! xAI released Grok 4.5!

Hell, even Meta are back in the game with a new model. Whaa?

Nice! More competition. Maximum boost!

We are drowning in state of the art models. Including Fable.

Fun if you still have usage. Less fun if you burned through your Fable allowance before the deadline and then Anthropic wandered back in like, "actually lads, have five more days."

Annoying.

But I reckon the bigger point is not Fable. Or Sol. Or Grok being "Opus-class" (we shall see, Elon, we shall see…).

The wrong question is: which model wins?

The useful question is: how do they work together?

The model is not the system

Models change. But you can’t be constantly switching or you’ll get NOTHING done.

Instead focus on your workflows. And make the model agnostic.

If your entire AI workflow is "open the smartest model and ask it everything", you are going to have a miserable few years. The smartest model will change - weekly. Or daily if this week is anything to go by! The price will change. The access rules will change.

Change is the only constant as some smart Greek who I can’t be bothered to look up said. ChatGPT would have known…

This week it is Fable, Sol, Grok and (checks notes) Meta Spark 1.1. A year from now it will be other names.

So the durable skill is not brand loyalty. It is workflows that can swap models in and out without collapsing. I wrote about this from the pricing side in the Fable is going away issue. Frontier intelligence is getting rationed. Sometimes by money. Sometimes by policy. Sometimes by capacity.

So the smart play is to use the cleverest model where it actually matters. Use the cheaper thing where it is good enough. Connect the two.

Use expensive intelligence where it matters

The simple split is planner model and worker model. Or an orchestrator and its agents if you fancy.

Planner model: strategy, architecture, judgement, trade-offs, review. This is where you use Fable, Sol or whatever the current expensive biggest brain is.

Worker model: file edits, tests, implementation, cleanup, repeatable tasks. This can be Codex, a cheaper GPT model, a hosted open model, a local model. Whatever clears the bar.

On our live AI Hour call last night Adam talked about how he recently used Fable to scope out a project to convert a PHP project he made 20 years ago into a more modern form. Fable got to work on the strategy and the delegated step by step work to Opus and Sonnet. A big chunky project that took the smaller AIs 4 days(!) but meant saving potentially thousands one tokens by intelligently using the right model for the job. Super valuable skill.

ClaudeDevs had a good version of this on X: use Fable as the advisor, then let a cheaper executor call it when it needs guidance.

As I mentioned in the last newsletter I spent about 900 million tokens on Codex on one heavy day (June 14th to be precise). If I ran that through Fable API pricing, the back-of-the-envelope number was around ~$15,000. For one day.

This is why we need to get smarter with how we use different tools.

I personally do not want to remortgage my house because I asked the clever model to fix a typo in a test file.

So do not use the smartest model just because you have access to it. Paradoxically that is NOT a smart thing to do.

But HOW?

Put the tools in the same room

The practical bit is much less mystical than people make it sound.

You need a shared folder.

For most of this work, that means a GitHub repo. If that phrase makes you feel itchy, think of it as Google Drive with version history and fewer vibes. I’ve talked about how to set up Github in this Github 101 guide.

Could you use Google Drive? Technically? But I wouldn't.

GitHub is better when multiple agents are making changes because it can track what changed and what needs merging. Google Drive is lovely for documents. It is a bit crap for agent handoffs. Because they’ll all be in there working on the same files and making a mess of it all.

This is the same reason I keep talking about building an AI brain / shared vault. A shared workspace for ALL your AI tools. (Also the topic of our AI Hour live call last night).

Inside the repo you add boring text files:

• README.md

• AGENTS.md

• CLAUDE.md

• tasks.md

• decisions.md

(By the way MD just means Markdown. Basically a stripped-down text file. Less scary than it sounds)

Tell Claude/Fable to write the plan and put the decisions in the repo. Then tell Codex to pick up the work and implement it. That’s (at base) it.

Start manual. Add machinery later.

As well as having a shared workspace we can also directly connect out tools to one another in various ways.

There are three ways to connect the tools.

First: manual handoff. Open both tools from the same repo. Claude writes the plan. Codex does the work. Claude reviews. You are the messenger. That works. It is also a bit of a faff because you become the bottleneck.

Second: plugins. Claude Code can call Codex. Codex can call Claude Code. Both have plugins that you can install in the other. Or if you are using Cursor then install both!

Third: MCP, CLI or custom wiring. This is where the tools talk directly under the hood. This is the most “advanced” but bizarrely probably also the easiest. You can literally ask the AI:

I want you to connect this project to Codex so you can 
hand implementation work off to it. 
Work out the best CLI or MCP route and set it up.

That is it. Seriously. The AI is smart. It’ll work the rest out.

My beginner loop

This is the version I would use if you are starting from scratch today. This WILL change. Might not even be the same a month from now. But the principal holds.

Start in the best planner. For me, right now, that is Fable. I give it the messy brief and tell it to interview me. Then I switch to voice and yap into the microphone.

Fable turns that into a project plan, file structure and shared instruction files.

Then Codex implements. It has a LOT more usage so can work on tasks doggedly for long horizons without bankrupting you.

Then the work goes back up to Fable or another stronger reviewer. It checks the result and sends fixes back down.

That loop itself matters more than whether the model name on the box is Fable, Sol, Terra, Luna, Grok or whatever else gets announced while I am trying to have breakfast….

Do not marry models

The warning is simple:

Do not just use the smartest model because you can.

That’s dumb(!).

Instead use the best model for the job at hand.

One plans. One writes. One reviews.

Remember that the. model names will change. The limits will change. The prices will change.

Your job is not to predict which logo wins next week. Leave that to the Twitter bros.

Your job is to build the loop so each release does not matter.

To the Task,

Kyle