ChatGPT Fails at Coupon Math? Open-Source Tool Cuts McDonald's Price in Half

When AI Meets Coupons: ChatGPT's Math Weakness

Asking ChatGPT to calculate McDonald's coupon combinations might end up costing you more money. This isn't a joke—large language models are unreliable at deterministic computation tasks, especially when stacking coupons, spending thresholds, and discount vouchers with complex rules, where they tend to hallucinate or make calculation errors. A recent YouTube channel showcased an open-source tool called "MCD Saver Strategist" built precisely for this pain point. According to the video, ChatGPT repeatedly fails at coupon math, while this tool connects an AI agent to McDonald's real-time ordering data through deterministic code to directly compute the lowest price.

LLM "Hallucinations" and Arithmetic Deficiencies: Large language models (like ChatGPT) are essentially probability-based text generators—their outputs aren't derived through deterministic calculation but by predicting the next word based on patterns in training data. This mechanism makes them prone to "hallucinations" (generating plausible-sounding but actually incorrect content) when handling precise arithmetic, logical reasoning, or rule combinations. Particularly with multi-step calculations (such as stacking coupons, spending thresholds, and discounts), the model may miss constraints or confuse the order of operations, producing incorrect results. This isn't because the model is "not smart enough"—it's that its architecture isn't suited for deterministic computation. Therefore, in scenarios requiring precision, the industry typically combines LLMs with external computation tools (such as calculators and rule engines), letting the LLM handle understanding and orchestration while delegating calculation to traditional code.

How the Tool Works: From "Guessing Discounts" to "Computing Discounts"

The core idea of MCD Saver Strategist is simple: don't let AI "guess" discounts—let AI orchestrate deterministic code logic. As mentioned in the video, the tool runs deterministic code over real-time coupon stacking, iterating through all possible combinations to find the lowest price. The test subject was a Big Mac meal, and the final checkout price was cut by 53%—essentially half price.

This design cleverly separates the LLM's semantic understanding from the rule engine's precise calculation: AI handles understanding user intent and fetching data, while the math is left to traditional code. This avoids the LLM's arithmetic uncertainty and improves the credibility of results.

Deterministic Code and Rule Engines: Deterministic code refers to code logic that produces the same output given the same input, no matter how many times it runs. In coupon combination scenarios, deterministic code enumerates all possible coupon usage orders and combinations, strictly calculating the final price according to rules, avoiding any randomness or approximation. A rule engine is a software system specifically designed to execute a set of business rules, separating rules from code for easier management and updates. MCD Saver Strategist uses deterministic code to iterate through all combinations, ensuring it finds the true lowest price rather than relying on the LLM's "guesses." This hybrid architecture of "AI orchestration + deterministic computation" is becoming increasingly common in tasks requiring high reliability.

Real-World Barriers: Regional Lock and Developer Configuration

The video author also acknowledged the tool's limitations. First, it's currently only available in the Chinese market, limiting users in other regions globally. Second, the tool requires developer-level configuration, making it difficult for ordinary consumers to use directly. The author categorized it as "underweight"—stunning concept, but region-locked.

Takeaway: The "Deterministic Computation" Pattern for AI Agents

Although MCD Saver Strategist itself is still rough, it points to an important direction: AI agents shouldn't handle all computation themselves but should serve as orchestrators of deterministic tools. In scenarios involving prices, inventory, and discount rules, combining LLM reasoning with precise code execution may be a more reliable architecture. For developers, this "AI + rule engine" approach can be reused in e-commerce price comparison, travel booking, food delivery discounts, and other domains.

The video author ended by inviting viewers to comment "test" to get the link, showing community interest in such tools. However, until regional restrictions and configuration barriers are resolved, it's more of a proof of concept than a mass-market product.

AI Agent Architecture: An AI agent refers to an intelligent system capable of perceiving its environment, making autonomous decisions, and executing actions. In LLM-driven agents, the model acts as the "brain," responsible for understanding user intent, planning steps, and invoking tools. However, an agent's reliability depends on whether the tools it invokes are reliable. Combining LLMs with deterministic tools (such as calculators, database queries, and rule engines) to form a "tool calling" pattern can significantly improve agent performance on complex tasks. For example, in price comparison scenarios, an agent can first parse user needs, then call price APIs to fetch data, and finally use deterministic algorithms to compare prices. This architecture leverages the LLM's semantic understanding strengths while avoiding its computational weaknesses—an important trend in current AI applications.

Conclusion

ChatGPT's failure at coupon math reveals the boundaries of LLMs, and the open-source tool MCD Saver Strategist fills this gap with deterministic code. Although it's currently limited to China and requires developer configuration, its design philosophy is worth noting: let AI do what it's good at, and leave computation to code.