01What Kilo Code Fable 5 Is and How It Works
Kilo Code Fable 5 is a method that uses a frontier model to plan and a cheaper model to build, achieving the same output for less cost. In this episode, Julian explains how Kilo Code's case study shows that a detailed plan lets any model build the identical service.
Fable 5 was the most powerful AI ever made public, with an 80% score on SWE bench pro. But it was shut down worldwide three days after launch. Kilo Code wanted to prove that you don't need Fable 5 itself to get that level of intelligence.
The key is planning. Two frontier models planned the same service, then both built the winning plan from identical starts. The resulting services were identical down to which individual users had a 35% rollout enabled.
Both passed 15 checks. The only difference was token usage and cost.
02Main Features of the Kilo Code Planning Method
The main feature is separating planning from building. You use a smart model to write a plan, then switch to a cheaper model to build it. This works because a pinned plan decides every fork, leaving nothing to guess.
- Planning with a frontier model: use your most powerful model to create a detailed plan.
- Building with a cheaper model: use a free or low-cost model to implement the plan.
- Review agents: check the build for bugs and security and logic issues before shipping.
- Reusable asset: one plan can be fanned out to multiple cheaper models to create variations.
- Model flexibility: if a model gets taken down, you swap the builder and keep moving.
This approach reduces token usage and cost while keeping quality. According to Kilo's research, you can get identical outputs for 59% less cost, which is a 2.4x gap that scales as you grow.
03How to Set Up Kilo Code Fable 5 Planning
Setting up the Kilo Code Fable 5 method inside Kilo is simple. You need two prompts: one for the planner and one for the builder. Here are the steps Julian shows in the episode.
- Open Kilo Code in your editor.
- Select a frontier model like Opus 4.8 for the planning phase.
- Prompt it to write a very detailed plan, covering every decision.
- Once the plan is ready, switch the model to a cheaper one like Kimikator 17 or a free model like Nemetron 3 Ultra.
- Use the same plan for the build phase.
- Run review agents to check for bugs and security issues.
- Ship the build.
You can also build a skill around this in Claude, telling it to use a specific model for building and a powerful model for planning. The process is designed to be easy for beginners.
04How to Use the Kilo Code Fable 5 Method in Practice
In the episode, Julian walks through a real example. He says to Opus 4.8: "Write a very detailed plan." Then he lets a cheaper model build it. He also mentions using free models like Nemetron 3 Ultra or Minimax M3 for the build phase.
The old way was to point every job at the most powerful model, which uses more tokens and costs more as you scale. The new way uses a genius AI for planning and cheaper models for building, with identical output.
You can also use this method with Fusion, a different API that claims Fable-level intelligence. Fusion uses multiple agents to come up with a plan, then a judge fuses the outputs into one answer. You can have Opus, Gemini, and Grok working together, and on benchmarks it potentially hits frontier levels.
Julian emphasizes that you should test it yourself rather than trusting benchmarks. He shows how you can run the whole system in one tool with two prompts.
05Results and Comparisons from Kilo Code's Case Study
Kilo Code ran a case study and published the results. They had two frontier models plan the same service, then made both build the winning plan from identical starts. Both services passed all 15 acceptance checks and were identical down to the 35% rollout detail.
The cost difference was significant. Planning with Claude Fable 5 and implementing with GPT 5.5 produced the same service for 59% less than using Claude Fable for both phases. The cost gap is 2.4x, which scales to a lot as you grow.
Julian also compares the old way and the new way. The old way uses more tokens and is less efficient. The new way uses less tokens overall and is more efficient.
He says that if you have a complete plan that leaves nothing to guess, both models build the exact same plan and pass all checks.
He cautions against paying too much attention to benchmarks. Instead, he recommends testing yourself to see what works for you.
06Tips, Limits, and Mistakes to Avoid
Julian shares several tips and warns about common mistakes. First, don't wait for one model to clearly win, because Fable 5 was the best and then it got removed. You can use the planning method now.
Second, some people think the best model should do everything, but that's not necessary. The best model is great for judgment and planning, but you don't need it to do everything if you have a fail-safe plan.
Third, a vague plan leaves forks open, so each model will guess and produce different apps. A pinned plan decides every fork, so there's nothing left to guess, and any model can land on the same build.
- Test with free models like Nemetron 3 Ultra, local models, or free APIs like Alpha.
- Use review agents to catch bugs and security issues before shipping.
- Remember that the plan is reusable, so you can build from it forever.
He also notes that this sounds technical, but you can do it with two prompts inside Kilo. Building a skill around it in Claude is another option.
07Who Kilo Code Fable 5 Is For and Why It Matters
Kilo Code Fable 5 is for anyone who builds with AI agents and wants to cut costs without losing quality. It matters because it lets you get frontier-level intelligence without relying on the most expensive models.
If a model you use gets taken down, you can swap builders and keep going. The plan survives, so you don't have to start over. This is a big advantage for teams and individuals who depend on AI for development.
Julian also mentions that the whole system is available inside the AI Profitable boardroom. He offers a full agent operating system for implementing AI agents working together. You can use a powerful model for planning and delegate to sub-agents for building.
Whether you're a beginner or an expert, testing this method yourself is the best way to see if it works for you.
