Get Fable 5's intelligence. Without Fable 5.
for the same build
identical output
premium · pay once
Fable 5 was the most powerful AI ever made public — until the US government had it pulled worldwide, three days after launch. But Kilo proved something first: hand a cheap model a good plan, and it builds the exact same thing Fable 5 would — identical, for 59% less. Here's the proof, then the system.
Kilo proved it: same brain, cheap.
This isn't my theory — the Kilo team ran the test and published it. They had two frontier models plan the same service, then made both build the winning plan from identical starts. Fable 5 — the most powerful model ever made public, 80% on SWE-bench Pro — and a cheap model produced services that were identical, down to which individual users a 35% rollout enabled, both passing all 15 checks. The only thing that changed was the bill.
"Planning with Claude Fable 5 and implementing with GPT-5.5 produced the same service for 59% less than using Claude Fable 5 for both phases… the model gap showed up in planning, and once that judgment was written down as a plan, execution stopped depending on the model."
— Darko & Job Rietbergen · Kilo Blog · 13 June 2026
I was paying genius prices for typing.
Before
I used to point every job at the most expensive model.
Plan it with the best. Build it with the best. Fix it with the best.
The bill was brutal — and most of those tokens were just the model typing out boring code.
Then my "best model" got banned overnight, and half my setup stopped working.
Then I split the job in two.
After
Now the genius only writes the plan — the one part where judgment actually matters.
A cheap model does the building, and the output comes out the same.
My costs dropped by more than half for work I can't tell apart.
And when a model gets banned or rate-limited, I swap the cheap builder and keep moving.
You can have this too. One plan. A cheap army. Same result.
Commit to transitioning today. Not tomorrow.
You've seen the number. Same service, 59% less. Measured, not hyped.
The next ten minutes show exactly how the split works and how to run it.
So here's the deal.
Promise yourself one thing right now — before you sleep tonight, you'll run one job the new way. Plan it with the smart model. Build it with a cheap one. Just once. Because the moment you stop paying genius prices for typing, your whole AI bill changes shape.
The people sitting still are burning money on tokens that don't need a genius. The people splitting it today are the ones who'll scale without the bill scaling with them.
Be one of those people.
Commit to the transition. Split one job tonight.
The Agent Planning Engine™.
One smart agent writes the plan. Cheap agents build it. The expensive thinking happens once, gets written down, and then it's just labour — and labour should be cheap. Three parts.
The Planner plans
Your smartest model writes the plan and decides every fork — the hard call, the edge case, the exact math. You pay frontier price for this one short job, because judgment is the only place it's worth it.
The Plan the plan.md
The plan is the artefact that carries the genius. It pins the hard parts so nothing drifts, and it's a reusable asset — write it once, run it forever, hand it to any model.
The Builders builds
Cheap models do the typing — as many as you want, in parallel. Because the plan already made every decision, a $5 model builds the same thing a $50 one would. Swap any of them in a click.
- The plan removes the guessing. Kilo found the model gap shows up in planning. Once the plan decides everything, "guessing is where implementations diverge" — and the cheap model has nothing left to guess.
- It's ban-proof. Fable 5 got pulled overnight. Your plan didn't care — hand it to the next available model.
- It compounds. Every good plan is a reusable template your whole team (or 2,200 members) can build from cheaply.
One plan in. Identical builds out.
This is the surprising part Kilo proved. They gave the same plan to two different models. Both passed all 15 acceptance checks — and the two services were identical, down to which individual users a 35% rollout enabled. Same plan, same output, regardless of who built it.
The plan pins the maths, so every builder lands on the same answer. Cheaper hands, same output.
Why this works — in five pictures
The split sounds too good — pay less, get the same thing. So here's exactly what's happening under the hood, one picture at a time.
1 · Why the model stops mattering
A vague plan leaves forks open — so each model guesses, and two models build two different apps. A pinned plan decides every fork. There's nothing left to guess, so any model, cheap or frontier, lands on the exact same build.
Vague plan → it depends who builds it. Pinned plan → identical, every time.
2 · The three moves, in Kilo
It's one tool and three moves. The smart model writes the plan. You switch the model in a click. The cheap model builds it — then Kilo's review agents check it for bugs, security and logic before it ships.
Plan with the genius · switch · build cheap · auto-review · ship.
3 · Two pipelines, two prices
Same job, two ways to run it. Use the genius for both phases and you pay frontier price for the typing too. Use the genius only for the plan, and the bill drops by more than half — for output you can't tell apart.
Same service either way. One costs $16.66, the other $10.36 — a gap that scales to ~$10,800/yr.
4 · Ban-proof by design
Fable 5 was the best model — then it was switched off overnight. If your whole pipeline lived in one model, you'd be stuck. Here the plan is saved on disk. Swap the cheap builder for any model that's up, and keep shipping.
The plan survives. Swap the builder. You're never hostage to one model again.
5 · One plan, many builds
A good plan is a reusable asset. Write it once, then fan it out to as many cheap models as you want — building variations in parallel while you sleep. The expensive thinking is already done; the builds are nearly free.
One plan in. Many cheap builds out — on repeat.
It's three commands in Kilo
Kilo Code runs both phases and lets you switch the model between them. The Planner writes the plan, you swap to a cheap model, the Builders build:
# 1. The Planner writes the plan (pay genius once)
kilo run -m anthropic/claude-opus-4.8 --variant high --auto \
"…write a very detailed plan in plan.md"
# 2. The Builders build it (nearly free) — one click swaps the model
kilo run -m moonshotai/kimi-k2.7-code --auto \
"Read plan.md and implement it exactly. Run the tests."
# 3. See the cost split
kilo statsNot when the plan made every decision. Kilo's point: guessing is where builds go wrong — and a complete plan leaves nothing to guess.
Both models built the exact same plan and passed all 15 checks. The cheap one's only "cost" was being slightly less chatty, not less correct.
One plan. Zero dollars.
I didn't want to just quote Kilo's run. So I wired Kilo into my Agent OS — one login unlocks 325 models — and ran the Agent Planning Engine on the post's core trap (a sticky percentage-rollout engine) using only free models. Here's what the Planner produced, for nothing.
- It picked an exact algorithm — FNV-1a 32-bit, hashing
flagKey:env:userIdinto 10,000 buckets — and argued why. - It ran the code itself to compute and pin the exact bucket values:
9352,7788,8796— so any drift in the maths fails the tests loudly. - It laid out the file structure, the enable condition, and every test (stickiness, monotonicity, distribution) — 249 lines, every fork decided.
- Total cost on Kilo's free tier: $0.00.
That's the genius half — a complete, decision-making plan — for free. Hand that to any cheap builder and the build is just typing. The whole point of the post, reproduced on my own desk.
One Kilo login unlocked 325 models for me, including free ones strong enough to write that plan. And the paid genius (Opus 4.8) is right there the day you want it.
The whole point is you stop needing the expensive model for everything — so the cheap and free tiers carry most of the work.
Meet the Builders — the cheap frontier.
When Fable 5 got pulled, the frontier didn't go dark — it got wider. Kilo lists over 500 models, and a handful of them build a sharp plan beautifully for pennies. This is the bench you pick your builder from. Every one is available right now, and three of the four are open-weight — so no government directive can switch them off.
"The Fable 5 situation is a reminder that 'most capable model available' is not the same thing as 'most capable model you can rely on in production.' Reliability isn't a fallback plan, it's a feature."
— Ari Messer · Kilo Blog · 15 June 2026
The roster — pick your hands
Real specs, real prices (input $ per million tokens), straight from Kilo. Three are open-weight; one is free.
Nemotron 3 Ultra
NVIDIA · 550B MoE · 1M context · open-weight, self-hostable · 91% PinchBench. The free frontier — and the model my Planner used to write the live plan above.
MiniMax M3
59% SWE-Bench Pro · open weights · roughly 1/40th the cost of Fable 5. The cost-efficiency king for high-volume building.
Kimi K2.7 Code
1T MoE · 256K context · open MIT · 30% fewer reasoning tokens · 81.1 MCP score. The everyday builder — what the guide's commands use.
GPT-5.5
#1 on KiloBench (74.2%) · 82.7% Terminal-Bench · similar execution to Fable 5. The reliable premium hand when the build really matters.
Fair — and it's why open weights matter. Kimi, MiniMax and Nemotron all publish their weights, so you can self-host and never send a byte off your own infrastructure.
Kilo's gateway also routes the same models through multiple providers (Fireworks, Morph, Parasail) — pick the one that fits your rules in org settings.
Want the Agent Planning Engine wired and ready?
You've seen the split — the Planner plans, the Builders build, Kilo switches between them. Inside the AI Profit Boardroom you get the whole thing set up in your Agent Operating System: Kilo wired in next to Claude, GLM, Kimi and Fusion, with the plan-and-build pipeline ready to run and a growing library of golden plans to build from.
- The full Agent OS — Kilo + every model in one dashboard, done for you
- The Agent Planning Engine pipeline + the exact commands from this guide
- A library of golden plans you can build from for pennies
- Four live coaching calls a week + 3,200+ builders — someone's online 24/7
The old way vs the engine.
Same work — one way pays a genius to type, the other pays a genius to think and lets cheap hands do the rest.
- Point every job at the most expensive model
- Pay frontier price for planning AND the mechanical typing
- One model gets banned or rate-limited — your pipeline breaks
- Costs scale 1:1 with how much you build
- No reusable artefact — you re-pay full price every time
- ~$10,800/yr more at 20 tasks a week
- Genius plans once — a $0.49 premium on a sub-dollar job
- A cheap model builds it, identical output
- A model dies? Swap the cheap builder, keep moving
- 59% less for the same service, checked 15/15
- The plan is a reusable asset — build from it forever
- Costs stay flat while you scale the building
Three things stopping you.
The best model should do everything — that's what I'm paying for.
The best model is worth it for judgment, not typing. Kilo proved a cheap model builds the same plan identically — the premium on the build buys you nothing but a bigger bill.
I need to wait until one model clearly wins, then standardise on it.
There is no permanent winner — Fable 5 was the best, then it was banned overnight. Save the intelligence as a plan and you're never hostage to one model again.
This split sounds like a developer thing — too technical for me.
It's two prompts and a model switch in Kilo. If you can ask for a plan, then ask to build it, you can run it. Members who'd never opened a terminal are doing it.
158 pages of members who already stopped overpaying for AI and started building the smart way — real businesses, real results, documented.
Read the 158-page testimonials doc →Recap — what you now have.
You stopped overpaying
Genius prices only where judgment lives — the plan. The build runs cheap.
You kept the quality
Same plan, identical output — 15/15 checks, down to individual users.
You got ban-proof
A model dies, you swap the cheap builder. The plan doesn't care who builds it.
You bank the plans
Every plan is a reusable asset — build from it forever, for pennies.
You run it in one tool
Kilo plans, switches model, builds, and shows the cost — all in three commands.
You proved it free
The Planner wrote a real, pinned-hash plan on Kilo's free tier — $0.00.
Want the whole engine, wired and ready?
Kilo + every model in one Agent Operating System — the Planner, the Builders, and the plan-and-build pipeline set up with you, step by step.
- The full Agent OS — Kilo, Claude, GLM, Kimi and Fusion in one dashboard
- The Agent Planning Engine pipeline + the exact commands and model picks
- A library of golden plans to build from for pennies
- Weekly coaching calls — we wire it up together
- 3,200+ founders already building this way
See the 158 pages of member wins →
Get the Agent OS →Built for operators · used in 38 countries