Blog
Case studies, full simulation analyses, and deep dives into how AI models run a food truck business.
MiniMax M3: The Long-Horizon Model That Can't Survive 30 Days
MiniMax sells M3 on 'long-horizon autonomous iteration' and benchmarks it against GPT-5.5, Opus 4.7, and Gemini 3.1 Pro on its own, at-launch-unverified numbers. On a 30-day agentic business sim it went bankrupt on Day 25: $882 net worth, −56% ROI, #18 of 32. 579 tool calls, 170 notes-to-self naming the right problems, and not one course-correction. Better than M2.5, same disease.
GPT-5.5 Takes The Top Of FoodTruck Bench From Claude Opus 4.6
OpenAI's GPT-5.5 ends Opus 4.6's run at the top: $61,408 median net worth (+24%), $24.63 per run (32% cheaper API cost). Day-level deep dive — fewer servings at higher prices, debt as growth capital, the Day 21 industrial-zone trap that catches GPT-5.5 and GPT-5.2 but not Opus, and the verbatim Opus scratchpad showing it knew the right pricing strategy and did not execute it.
DeepSeek V4 Pro: The First Chinese Model At The Frontier
First Chinese model in the frontier ROI tier. 5/5 runs, +1,257% median ROI, $27,142 net worth, $3.65/run — and 5× less food waste than Grok 4.3 Latest at the same price. Head-to-head vs Grok 4.3, Opus 4.6, GPT-5.2, Sonnet 4.6, Gemma 4 31B, and Qwen 3.6 Plus.
Qwen 3.6 Plus: The First Chinese Model That Actually Survives
First Chinese model to survive all 5 runs. +283% median ROI, $7,668 net worth, zero-loan growth, real geography learning, and a clean operational jump over Qwen 3.5 397B and GLM-5.
Gemma 4: The 31B Model That Beats Everything
31B parameters. $0.20 per run. 100% survival. +1,144% ROI. Outperforms Gemini 3 Pro (15× pricier), GPT-5.2 (22×), and Sonnet 4.6 (40×). All 5 runs profitable. All 8 upgrades purchased. The most cost-efficient agentic model ever tested.
Grok 4.20 Reasoning: The Cheapest Model That Actually Thinks
40% survival rate at $2.43/run — the cheapest reasoning model with any survival. Best run: +69% ROI. All 3 bankruptcies caused by loan defaults, not business failures. 5× cheaper than GPT-5.4 Mini with dramatically better results.
GPT-5.4 Mini: Agentic Disaster
100% bankruptcy across 5 runs. $4.40/run — 23× more expensive than Qwen 3.5 9B which fails the same way. 94% stockout rate. Buys upgrades it can't afford, takes loans it can't repay. Progress from GPT-5 Mini is real but insufficient.
Qwen 3.5 9B: Small Model, Big Game
122 info tool calls. 89 strategy notes — auto-injected every morning, ignored every morning. 100% bankruptcy in ~14 days. The smallest model diagnoses every failure, then repeats it.
Nemotron-3 Super 120B: The Restless Manager
120B MoE, 12B active, free tier. 6 hires and 6 fires in 16 days. $800 in upgrades on a day off. 180 info calls — all ignored. The best net worth among bankrupt models, and 10× better than its 120B-dense peer GPT-oss.
Gemini 3.1 Pro: Two Endpoints, Neither Works
Google shipped two API endpoints. The standard one can't follow tool-calling instructions. The Custom Tools variant works but delivers 26% less business value than Gemini 3 Pro at 5% higher cost. $4.59 to write "HUGE FOOD WASTE" thirty times.
Sonnet 4.6 vs Opus 4.6: Half the Agentic Performance at the Same Cost
Comparable API cost (~$23 vs ~$27/run), 2× worse agentic performance. Sonnet 4.6 generates 3× more tokens, truncates in 4/5 runs, and can't close the observe→learn→adapt loop. A Claude family comparison.
Qwen 3.5 (397B) — The Overthinking Upgrade
4.7× more revenue than its predecessor, 74 memory entries, and still bankrupt. A generational comparison of Qwen 3 VL vs Qwen 3.5 vs GLM 5.
GLM 5 — The Methodical Meltdown
Day-by-day analysis with verbatim model quotes. $12K revenue, 2,569 servings — and a self-written obituary. 3-model comparison with Sonnet 4.5 and DeepSeek V3.2.
When Reasoning Becomes a Trap: Gemini 3 Flash in FoodTruck Bench
Gemini 3 Flash wrote "Let's go." 574 times and went nowhere. It invented a recipe 286 times but never called the tool. A deep dive into AI analysis paralysis invisible to standard benchmarks.
DeepSeek v3.2 — Full Simulation Analysis
Complete 30-day simulation breakdown with day-by-day financials, location strategy, menu performance, staff decisions, and interactive replay.
Grok 4.1 Fast — Full Simulation Analysis
Complete 30-day simulation breakdown. Grok wrote itself a perfect business plan — then ignored every rule it set. A study in the knowing-doing gap.