Yonathan Cohen Reveals Open-Source Skill to Slash AI Model Costs by 90%

In a recent LinkedIn post, Yonathan Cohen shares a cost-saving strategy for leveraging advanced AI models, demonstrating how a custom open-source ‘skill’ can reduce expenses by up to 90% while maintaining accuracy. Cohen highlights his experience with Fable 5, an AI model he praises as the best he has used, despite its significantly higher cost compared to models like Haiku.

Cohen’s core principle for this cost reduction is elegantly summarized: “The smart model thinks once. The cheap model executes a thousand times.” This philosophy underpins a three-command approach designed to optimize AI processing and cost.

The ‘Skill’ Architecture for Cost Optimization

The strategy, which Cohen refers to as a ‘skill,’ involves a sequence of commands to process information efficiently. According to Cohen, the process begins with the ‘distill’ command.

“Fable studies examples you already handled right and writes a reusable rulebook. You pay for the thinking once.”

As Yonathan Cohen explains, this initial step involves using a powerful, albeit expensive, model like Fable 5 to analyze successful past interactions and generate a set of rules or a knowledge base. The key benefit, Cohen emphasizes, is that the significant ‘thinking’ or analysis is performed only once during this distillation phase, amortizing the cost over many future uses.

Applying the Rulebook with a Cheaper Model

Following the distillation, the ‘classify’ command comes into play. This step utilizes a more cost-effective model, such as Haiku, to apply the generated rulebook to new incoming data.

“Haiku applies the rulebook to every new request. On my benchmark (15 labeled prospect replies): same accuracy as Fable, 93% vs 93%, at 10x lower cost.”

Cohen’s benchmark results, measured live on the API, show remarkable parity in accuracy between the expensive Fable model and the cheaper Haiku model when using the distilled rulebook. This demonstrates that the ‘thinking’ performed by Fable in the initial stage allows the simpler Haiku model to achieve comparable results at a fraction of the cost.

Analyzing Large Volumes of Data

For more complex tasks, such as analyzing lengthy meeting transcripts or long email threads, Cohen proposes an ‘analyze’ command. This part of the skill leverages Haiku to read and compress extensive information, while the more powerful model is reserved for specific decision-making based on concise briefs.

Cohen provides concrete figures to illustrate the financial impact of his method. He states that processing over 10,000 classifications using Fable alone would cost approximately $79. However, by implementing his ‘skill,’ the cost is reduced to around $8, representing a 90% saving.

Debunking a Common Costly Approach

Yonathan Cohen also cautions against a seemingly intuitive but ultimately more expensive approach: having the cheaper model constantly consult the more powerful one for every query. He notes that this method proved to be significantly more costly in his tests.

“The ‘obvious’ hack, Haiku consulting Fable while it answers, costs 19x MORE than Fable alone. Measured. The advisor thinks at full depth on every single call, and you pay for every token of it.”

This insight underscores Cohen’s argument that true cost efficiency comes from strategic application of powerful models for initial learning and rule creation, rather than continuous, token-by-token consultation for every task. The strategy, which Cohen mentions works within Claude Code, offers a practical blueprint for businesses looking to optimize their AI expenditures without compromising on performance.