This AI Trick Slashes GPT Bills By 70%

Your AI bills are too high. Not because AI itself is inherently expensive, but because you're likely over-provisioning your models. Many developers default to a single, powerful frontier model like GPT-5.6 Sol for every task, from planning to execution to review. This is akin to using a supercar for grocery runs – overkill, inefficient, and astronomically expensive.

AI inference costs now rank as the second-largest line item in many tech budgets. Ignoring this spiraling expense means leaving money on the table, critically impacting your bottom line and project viability. Optimization isn't optional; it's a critical, immediate need for every developer and business running AI workloads.

The real challenge isn't just cutting costs by switching to weaker, less capable models. That’s a false economy that sacrifices output quality. Instead, we must optimize workflows to retain the same, or even superior, final output while drastically reducing expenditure.

For example, a single task costing $81 with an exclusive GPT-5.6 Sol workflow screams inefficiency. This monolithic approach burns through tokens unnecessarily. We need a smarter strategy to tackle such problems, slashing costs without compromising the quality of the final artifact.

Model tiering, also known as AI model routing, represents the single highest-ROI architectural decision for managing AI costs. The core principle: stop indiscriminately using powerful, expensive frontier models for every step. Instead, allocate intelligence precisely where it's needed.

Adopt the Planner-Worker-Reviewer strategy for optimal efficiency. A Planner model, like GPT-5.6 Fable, initiates the process by architecting the entire solution. This ensures a high-level, intelligent blueprint before any heavy lifting begins.

Next, the Planner delegates specific tasks to multiple Worker models for execution. These are faster, cheaper workhorse models, such as GPT-5.6 Terra or Composer 2.5. They perform the bulk of the computational work—like generating extensive code—at a substantially lower token cost and much faster inference speed. Composer 2.5, for instance, offers significant cost advantages over higher-effort models.

Finally, a Reviewer, typically another frontier model like GPT-5.6 Sol, performs crucial quality assurance. It scrutinizes the Workers' output, ensuring alignment with the Planner's initial design and maintaining overall output quality.

This tiered resource allocation is why it works. Expensive, high-intelligence models handle complex thought and critical oversight. Cheaper, specialized models execute the routine, high-volume tasks. This method reduces task costs from $81 down to just over $25, making your AI workflows both faster and cheaper without compromising quality.

Relay, a free, open-source command-line tool, makes advanced AI model routing accessible to any developer. It's the traffic cop for your AI agents, abstracting away the complex orchestration of the Planner-Worker-Reviewer strategy. This intelligent router connects your preferred coding agents to a diverse ecosystem of AI providers.

Hook up tools like the ChatGPT desktop app (in Codex mode) or Gemini CLI. Relay then delegates tasks to a vast array of backend models, including:

• Groq, Mistral, Nvidia, DeepSeek
• Custom OpenAI/Anthropic endpoints
• OpenCode Zen/Go, Google Vertex AI
• Local models via Ollama or LM Studio

Relay lets you leverage powerful frontier models like GPT 5.6 Fable for intricate planning and architecting. Then, it hands off execution to cheaper, faster workhorse models such as Composer 2.5 or GPT 5.6 Terra for actual code writing. Composer 2.5, for instance, costs $0.50/M input and $2.50/M output tokens.

Finally, a dedicated frontier model like GPT 5.6 Sol can review the output, ensuring quality. This practical enabler makes model tiering easy to implement, slashing costs significantly; one task dropped from $81 to just over $25. For full documentation and setup, look at the project's GitHub: jacob-bd/relay-ai: Relay any model into any coding agent - GitHub.

The $81 problem? Solved. With Relay, a coding task that once cost $81 using a singular frontier model now costs just over $25. This isn't theoretical; the video's case study demonstrates a ~69% cost reduction by strategically routing prompts. Specifically, Fable handled planning, Composer 2.5 executed the code writing, and GPT 5.6 Sol performed the final review.

Beyond the stark cost savings, speed is a critical secondary benefit. Workhorse models—like Composer 2.5 for execution or GPT 5.6 Terra—are purpose-built for efficiency. They process tokens significantly faster than their frontier brethren for high-volume, repetitive tasks. This means faster code generation and quicker overall task completion, directly boosting developer throughput and accelerating project timelines.

This isn't merely a trick; it's AI FinOps made accessible. Implementing a Planner-Worker-Reviewer strategy via Relay offers the single highest ROI architectural decision for sustainable AI applications. You build faster, cheaper, and with predictable expenses, moving beyond the "throw money at it" phase of early AI adoption.

For any serious builder, granular cost control is non-negotiable. Stop burning cash on overkill models for every single step. Adopt tiered execution. Configure your routing, specify your models for planning, execution, and then review. Watch your bills shrink and your efficiency soar; this is how you build sustainable, scalable AI without breaking the bank.

What is AI model tiering?

AI model tiering is a cost-optimization strategy where you use different AI models for different stages of a single task. It involves using powerful, expensive models for complex work like planning and review, and faster, cheaper models for execution, like writing code.

What is the Relay AI tool?

Relay is a free, open-source command-line tool that acts as a router for AI models. It allows you to automatically send different parts of a task to different AI providers (like OpenAI, Groq, or local models via Ollama) to implement the model tiering strategy.

How much can model tiering save on AI costs?

The savings can be significant, often between 30-70%. The video example showed a specific coding task's cost dropping from $81 to just $25, a reduction of nearly 70%, by using this method.

Does using cheaper AI models reduce code quality?

Not with this strategy. Because a high-capability frontier model is still used for the crucial planning and final review stages, the final code quality should be just as good as if you used the expensive model for the entire process.