AI Strategy

What Using Fable 5 Taught Me About Return on Token Investment

The short version

  • Strategy is everything with AI. Not the tool, not the prompt. Fable 5 is vision-led, so it works backward from an end point, and naming that end point is the real work.
  • Spend cheap tokens first. Lock the vision on the leaner models (Haiku 4.5, Sonnet 5, Opus 4.8), then run Fable 5 clean on a full tank.
  • The metric that matters isn't usage, it's aim. Return on Token Investment: the people who win with AI aim it best, they don't use it most.
  • Updated September 29, 2026: the same capability is getting cheaper by about a third a month while the top model costs more. Watch the cost of a result, not the price of a token. Read the update.

I spent 75% of my time with Claude's Fable 5 not using it. That was the plan, and here's why.

Strategy is everything with AI. Not the tool, not the prompt. The strategy. Fable 5 isn't tactics-led, it's vision-led. It works backward from an end point, the way real transformation does.

And here's the thing. "Start with the goal" sounds easy, everyone says it. But if you don't know your end point, you're sunk, and if you wing it, you're just going fast to nowhere.

The hard truth is that naming the real vision is the hardest part, and it's only going to matter more as AI adoption deepens. Everyone can write a prompt. Few can define the roadmap that the prompt is meant to serve.

Through my strategy framework, I locked the endpoint before I ever opened Fable 5. I ran it first on the leaner models, Haiku 4.5, Sonnet 5, and Opus 4.8, each chosen for its strengths and its token rate. Cheap tokens up front bought a clear vision and a full tank to run Fable 5 clean.

From there, it ran like an assembly line. I set the goals and boundaries, Opus 4.8 shaped each prompt from the Fable 5 documentation, and Fable 5 was built inside the right project with updated working memory. I reviewed every output and passed my feedback back to Opus for the next pass. Around and around it went.

In 8 hours

Every one of these exists because the strategy came first.

  • A live MCP connector running a subscription (a small SaaS)
  • A financial model
  • A full brand and website
  • A legal foundation
  • A marketing engine
  • An operations app

They exist because the strategy came first.

The metric worth stealing

ROTI, Return on Time Invested, comes from lean and agile. I'm changing one word: Return on Token Investment.

Since writing this I built the calculator version of it, so you can stop estimating. Price your own prompt in the ROMTI simulator: your real prompt, your model, and the costs that never appear on the pricing page.

The people who win with AI won't be the ones who use it most. They'll be the ones who aim it best.

Strategy is everything with AI. The tool doesn't transform anything. A vision does.

That's the work I do, and I'd love to help you find your own R.O.T.I. 📈 Now it's your turn: what did you build with your Fable 5 sample, and what did it teach you?

Update · 29 September 2026

What two months of price changes taught me

I wrote this post on July 31. Since then OpenAI shipped a new model family, and the prices moved enough to teach a second lesson.

The model you already use got cheaper, fast. OpenAI's Sol model launched on July 9 at $5 per million input tokens and $30 per million output tokens. GPT-6 Sol arrived on September 22 at $2 and $10. That is 60% off input and 67% off output in 75 days, roughly a third off every month.2

The model on top costs more. GPT-6 Astra, OpenAI's new most capable model, is $10 and $50, five times GPT-6 Sol.3 On a $20 Plus plan it also buys fewer turns. OpenAI lists 5 to 45 Astra messages per five-hour window, against 15 to 150 for Sol.4

So the rate is the wrong number to watch. Take one real task: 100,000 tokens of context in, 10,000 tokens out.

  • GPT-5.6 SolJuly launch price$0.80
  • GPT-6 Soltoday$0.30
  • GPT-6 Astratoday$1.50
  • GPT-6 Lunatoday, the light model$0.015

Astra costs $1.20 more per run than Sol. If your time is worth $60 an hour, that is 72 seconds. If Astra saves you one extra round of review, it paid for itself. If it saves you nothing, you paid five times for the same answer. The token rate cannot tell you which. Only the result can.

Total cost of ownership is the real bill. It is the token rate, times the runs it takes to get a usable output, plus the minutes you spend checking it, plus the cost of switching when a model retires. That last cost is already here. OpenAI retired GPT-5.4 from Codex on August 31 and a Codex coding model on September 14, and GPT-5.5 leaves ChatGPT and Codex on October 14.5 That is three forced moves in seven weeks, and every saved prompt, default and automation pointed at those models has to be retested.

The lesson

The same capability is getting cheaper by roughly a third a month. Do not build your process around a price. Build it around a result, count what each result costs you in runs and minutes, and re-price every quarter. The cheapest model is the one you only have to run once.

Want your own numbers? Run them in the ROMTI simulator, now carrying GPT-6 Astra. Or see how we price AI per result for a whole business.

Disclosure. Prices are OpenAI's published list rates, checked September 29, 2026. One price cut is a data point, not a trend. The sample task and the $60 hour are my illustrations, not measurements. Prickled takes no commission, referral fee or incentive from any vendor named here.

Questions people ask about this

Straight answers, in case you skimmed.

What is Return on Token Investment?

It's a spin on ROTI, Return on Time Invested, a metric borrowed from lean and agile. Changing one word to Return on Token Investment reframes the goal: the win isn't using AI the most, it's aiming it best, so every token you spend serves a vision you've already defined.

Why run cheaper Claude models before Fable 5?

Because strategy should be locked before you spend on the heaviest model. Running Haiku 4.5, Sonnet 5, and Opus 4.8 first, each chosen for its strengths and token rate, buys a clear vision cheaply and leaves a full tank to run Fable 5 clean.

Does strategy really matter more than the tool with AI?

Yes. Everyone can write a prompt, but few can define the roadmap the prompt is meant to serve. The tool doesn't transform anything. A vision does, and naming that real vision is the hardest and most valuable part.

Is the cheapest AI model always the best value?

No. Price per token is falling fast, but price per result is what you pay. Total cost of ownership is the token rate, times the runs it takes to get a usable output, plus the minutes you spend checking it, plus the cost of switching when a model retires. A model that costs five times more per run is cheaper if it saves you one round of review.

How fast are AI model prices changing?

Fast. OpenAI's Sol model launched on July 9, 2026 at $5 input and $30 output per million tokens, and GPT-6 Sol arrived on September 22 at $2 and $10, about a third cheaper every month. At the same time the newest top model, GPT-6 Astra, launched at $10 and $50. Re-price your AI stack every quarter instead of locking a process to one price.

Sources & Notes

  1. ROTI (Return on Time Invested) is a feedback metric from lean and agile practice, most often used to rate whether a meeting or retrospective was worth the time spent. "Return on Token Investment" is my adaptation of it.
  2. GPT-5.6 Sol launched July 9, 2026 at $5 input and $30 output per million tokens (DataNorth). GPT-6 Sol launched September 22, 2026 at $2 and $10 (OpenAI). OpenAI's pricing page now lists GPT-5.6 Sol at $4 and $20.
  3. GPT-6 Astra: $10 input, $1 cached input, $50 output per million tokens (OpenAI model page).
  4. Messages per five-hour window on Plus (ChatGPT pricing documentation).
  5. Retirement dates (ChatGPT models documentation).