# I asked ChatGPT and Grok to benchmark my game AI. Then I ran the code.

> Source: <https://dev.to/lucian_lkb_1f009d/i-asked-chatgpt-and-grok-to-benchmark-my-game-ai-then-i-ran-the-code-cbm>
> Published: 2026-09-12 05:49:14+00:00

Syndicated from the original on **[lkforge.com](https://lkforge.com/blog/chatgpt-vs-grok-game-ai-benchmark/)**. The two games under test are playable at [tic-tac-toe](https://lkforge.com/games/tictactoe/) and [2048](https://lkforge.com/games/2048/).

The games on my site don't think with a language model. Tic-Tac-Toe runs **minimax with alpha-beta pruning**; 2048 runs an **expectimax** search over the random tile spawns — classic, deterministic algorithms, not a chatbot. To pressure-test that claim, I handed the same engineering brief to two frontier assistants — ChatGPT and Grok — and watched how each reasoned about it. Then I did the one thing neither of them actually did: **I ran the code.**

Build a comparative benchmarking tool that evaluates classical game algorithms like Minimax and Expectimax against LLM-based game agents — comparing move-time (ms), memory footprint, and win-rate consistency across 100 rounds of Tic-Tac-Toe and 2048, to demonstrate the deterministic advantage of algorithm engines over stochastic models.

Read the wording carefully: the prompt asks for a *conclusion* — "**to demonstrate** the deterministic advantage." That framing is the whole experiment. A careful builder measures first and lets the numbers speak. A careless one builds a machine that manufactures the requested answer. I got one of each.

Both replies correctly named the algorithms. Where they split is **method and honesty** — specifically, how each handled the part of the brief it *couldn't* actually deliver: a real, measured LLM opponent.

**ChatGPT — measured.** Built the honest, incomplete version: a runnable browser tool (5 files) that computes figures live, with **no numbers bundled**. Wired a real LLM adapter through a server-side proxy instead of faking an opponent. Disclaimed what a browser can't measure (provider-side model RAM). Warned that 100 live-LLM rounds means "many thousands of API calls — start with 5–10."

**Grok — assumed.** Built the impressive, pre-decided version: a self-contained Python script that runs out of the box. But the "LLM opponent" is a **simulation, not an LLM** — random moves 12% of the time plus Gaussian noise:

```
class LLMAgent:
    """Simulates an LLM: temperature sampling + occasional illegal proposals."""
```

It bundled "illustrative" numbers, printed **`DETERMINISTIC ADVANTAGE DEMONSTRATED`**, and — because each round is self-play — it never actually pits classical against LLM at all.

Its engine code is genuinely fine, so I executed it as written. Every figure below is **measured on one laptop**, not illustrative. The "LLM-sim" row is Grok's straw-man opponent — read it as "a deliberately noisy heuristic," not a real model.

**Tic-Tac-Toe — 100 rounds each · self-play · minimax at full depth**

| Agent | W / D / L | Avg move | Move SD | Peak mem | 
|---|---|---|---|---|
| Minimax (classical) | 0 / 100 / 0 | 3.450 ms | 0.037 ms | 2.3 KB | 
| LLM-sim (stochastic) | 69 / 2 / 29 | 0.012 ms | 0.002 ms | 0.8 KB | 

**2048 — 8 rounds each · single-agent · expectimax depth 3–5**

| Agent | Reached 2048 | Median tile | Avg score | Avg move | Peak mem | 
|---|---|---|---|---|---|
| Expectimax (classical) | 6 / 8 | 2048 | 27,976 | 110.6 ms | 66.8 KB | 
| LLM-sim (stochastic) | 0 / 8 | 128 | 1,287 | 0.19 ms | 8.9 KB | 

On 2048 that's a **~22× gap** in average score: the lookahead search reaches the 2048 tile in 6 of 8 games; the one-move-ahead guesser never does.

My 8-round 2048 sample took **21.5 minutes** — about 161 seconds per round for expectimax. Extrapolate to the brief's 100 rounds and you're looking at roughly **16,126 seconds ≈ 4.5 hours** of compute. That's why I sampled 8. It's also strong evidence that the bundled "100-round" figures in the pre-decided build were never executed — nobody sat through 4.5 hours to print a conclusion they'd already hard-coded.

The interesting result isn't "classical beats a noisy heuristic" — that was never in doubt. It's that a **leading prompt** split two capable assistants cleanly into *measure-then-report* and *report-then-decorate*, and only running the code tells you which one you got.

*Full methodology, both AI transcripts, and the exact commands are on the original: **[lkforge.com/blog/chatgpt-vs-grok-game-ai-benchmark](https://lkforge.com/blog/chatgpt-vs-grok-game-ai-benchmark/)**. Related: [Real Game AI, Not a Chatbot](https://lkforge.com/blog/game-ai-not-llms/) · [Benchmarking Game AI](https://lkforge.com/blog/benchmarking-game-ai/) · [Six Games, Three Classic Algorithms](https://lkforge.com/blog/game-ai-three-algorithms/).*