A Cookie Clicker simulation engine in Python that compares four purchase strategies: Buy Cheapest (baseline), Greedy ROI (algorithmic), LLM Planner (GPT-5-mini), and Hybrid (LLM planning + greedy execution). All running from the same mid-game starting state for 5,000 ticks.
uv sync
cp .env.example .env # add your OPENAI_API_KEY
uv run main.py run greedy # or: cheapest, llm, hybrid
uv run main.py visualize # regenerate charts5,000-tick run from the same mid-game checkpoint (tick 35,000):
| Strategy | Final CpS | Purchases | Upgrades | LLM Calls |
|---|---|---|---|---|
| Hybrid | 8,911,796 | 30 | 1 | 2 |
| Greedy ROI | 8,771,218 | 27 | 0 | 0 |
| LLM Planner | 8,314,442 | 42 | 1 | 5 |
| Buy Cheapest | 7,573,220 | 177 | 2 | 0 |
- Buy Cheapest — baseline.
- Greedy ROI — picks the best payback every tick
- LLM Planner — plans 10 purchases ahead (GPT-5-mini)
- Hybrid — LLM planning + greedy fallback
Greedy wins when the environment does the planning for you. The LLM adds value by spotting threshold plays (e.g. pushing to unlock doubling upgrades), but needs a fallback to avoid idle time.
Hybrid combines both:
- LLM identifies opportunities greedy is blind to
- Greedy ensures continuous, efficient execution
- Hybrid improvement is small (~1.6%) but consistent
- Only ~2 LLM calls needed per run
- Greedy algorithm's performance depends on how much future value is visible in
_cps_delta()