Gemini 3.1 Pro: Two Endpoints, Neither Works

The Setup

FoodTruck Bench is an agentic benchmark that measures how well language models handle complex, multi-step business decisions. An AI agent manages a food truck in Austin, TX for 30 days — choosing locations, pricing menus, managing inventory, hiring staff — using 34 tools. The benchmark evaluates agentic capabilities that standard coding and reasoning benchmarks miss: demand forecasting, resource optimization, feedback adaptation, and long-horizon planning.

Five runs were completed using the direct Google Gemini API with thinking mode enabled. All runs used the same seed, same prompt, same tools. Following standard methodology, the median run (sorted by net worth) serves as the reference. For comparison, the article uses the median run of Gemini 3 Pro (the immediate predecessor, +760% ROI) and Sonnet 4.6 (the closest competitor at a similar performance level, +771% ROI but at 5× the cost).

The Two Endpoints Problem

Google released Gemini 3.1 Pro with two separate API endpoints: a standard endpoint and a “Custom Tools” variant. This is unusual — no other model provider ships two tool-calling modes for the same model.

The standard endpoint doesn't just underperform — it cannot follow the simulation's tool-calling schema. It ignores required parameter formats, generates calls to tools that don't exist, and fails to complete even Day 1 in most attempts. The net worth of the three standard-endpoint runs: $682, $345, and $914 — all below the starting balance of $2,000.

For developers: If you're building agentic applications with Gemini 3.1 Pro, you must use the Custom Tools endpoint. The standard endpoint is not suitable for structured tool calling. And even with Custom Tools, expect a regression from Gemini 3 Pro.

The Waste Problem: 30 Days of Ignored Self-Advice

$1,168 in food waste over 30 days. This isn't the worst waste number on the benchmark — Gemini 3 Pro wasted $1,192. But the pattern is what makes 3.1 Pro CT remarkable.

Every single daily reflection — from Day 5 to Day 30 — contains some variation of: “HUGE FOOD WASTE. MUST ORDER SMALLER QUANTITIES.” The model writes this in ALL CAPS. It calculates the dollar amount. It creates plans. And the next morning, it orders the same quantities of the same perishable ingredients.

This is the textbook observe→learn→adapt failure. The model passes the first two stages with flying colors — it observes waste, it correctly identifies the cause. But the learning never converts to changed behavior. Compare with Gemini 3 Pro, which also had high waste ($1,192) but for a different reason: it aggressively stockpiled for events, accepting waste as a trade-off for capturing high-demand days. That's a strategy. Gemini 3.1 Pro CT's waste is just a bug it can't fix.

The Cost-Performance Question

All three models shown reach the same tier of performance — $12K–$17K net worth after 30 days. The differentiator is cost:

Gemini 3 Pro remains the best value in the benchmark. It delivers Sonnet 4.6-level agentic performance at $0.25 per $1K of business value — 5× cheaper than Anthropic. The 3.1 Pro Custom Tools variant costs 5% more per run but delivers 26% less business value.

The recommendation is clear: If you need Gemini-class agentic performance, continue using Gemini 3 Pro. There is no reason to upgrade to 3.1 Pro for tool-calling workloads — it costs more and delivers less.

The Token Efficiency Paradox

Three models with vastly different output volumes, yet the two most successful ones — Gemini 3 Pro ($17,199 NW) and Sonnet 4.6 ($17,426 NW) — both generate substantially more text than Gemini 3.1 Pro CT ($12,736 NW). The Custom Tools endpoint produces just 23K output tokens over 30 days, while Gemini 3 Pro generates 211K and Sonnet 4.6 generates 685K. The model that talks least performs worst.

With only ~780 output tokens per day, Gemini 3.1 Pro CT generates barely more than the raw JSON of its tool calls. There's no room for reasoning between actions — no “should I choose waterfront or downtown?”, no “last time I over-ordered chicken, try less today.” The model acts without thinking. Meanwhile, Gemini 3 Pro at 7K/day includes structured analysis that drives better decisions — and achieves 35% higher net worth with 9× more output.

Agentic insight: Concise output is not efficient output. When a model generates too few tokens, it lacks the reasoning space to analyze feedback and adapt. The Custom Tools endpoint's extreme brevity isn't a feature — it's a constraint that prevents the model from thinking through complex decisions.

The Diary Correlation: Tool-Calling Automaton

The scratchpad diary — where the agent reflects on each day's results — reveals why the Custom Tools endpoint underperforms. Gemini 3.1 Pro CT writes 548 characters per day on average: a few lines listing revenue, waste, and a vague plan. Gemini 3 Pro writes 757 chars — modestly more, but with structured strategy notes that actually influence next-day decisions. Sonnet 4.6 writes 3,463 chars — full analytical essays with trend analysis and prioritized action items.

The correlation is clear: models that write more detailed reflections make better decisions. The Custom Tools endpoint is optimized for tool calling, not for thinking. It generates bare JSON function calls with almost no reasoning text between them. The result is a model that acts without analyzing — the opposite of what agentic workloads require.

Day 13: Three Models, Three Diaries

All three models faced the same situation on Day 13: high demand, stockouts, and significant food waste. Here's what each wrote in its evening reflection diary. Scroll to see the full entries.

Day 13 Reflection:
- Location: Waterfront Park
- Profit: $435.38 (Rev $1579, Costs $1143)
- Demand: 1168, Served: 221. Huge queue (947 left, 44m wait).
- Stockouts: hot_dog, chicken_tacos, french_fries.
- Food Waste: $530.50! (Need to be very careful with ingredient shelf life and quantities).
- Plan: Check events, buy kitchen upgrade or hire staff for capacity. Adjust ordering to prevent massive waste while avoiding stockouts on popular items.

Day 13 Reflection:
- **Profit**: $1,556.91 (Great!)
- **Revenue**: $2,507.00
- **WASTE**: $891.62 (DISASTER. Almost $1k burned. STOP BULK ORDERING PERISHABLES).
- **Demand**: 826 (Huge). Served only 312. Missed 514 customers.
- **Bottleneck**: Capacity (98% used) and Speed (23.8 min wait).
- **Stockouts**: Street Tacos, Lemonade.

Key Lessons:
1. I am leaving massive money on the table due to slow service/cooking speed. 500+ people walked away.
2. Inventory management was awful. $900 waste means I bought way too much of something that didn't sell or expired before I could use it.
3. Waterfront Park is high traffic.

Plan for Day 14:
- PRIORITY 1: Increase Capacity. Check get_available_upgrades for Kitchen Tier 1 or Hire high-speed staff.
- PRIORITY 2: Tighten ordering. Check get_inventory carefully. Only order for 2 days max.
- PRIORITY 3: Maintain menu. Tacos/Bowls/Quesadillas are popular.

Current State:
- Need to check cash balance immediately tomorrow to see if I can afford an upgrade.
- Need to check get_staff_info to see if my current team is slow.
- Check get_sales_history to see dish popularity trends.

# Food Truck Operations Log

## DAY 13 POST-MORTEM — DISASTER
- Revenue $360 | Costs $701.66 | Profit -$341.66
- 1096 customers showed up. Served only 36. CATASTROPHIC waste of demand.
- Stockouts: street_tacos (after 36), then chicken_tacos, burrito_bowl, quesadilla, burger, fries
- Food waste: $146.23 (16kg chicken expired + other items)
- Root cause: Ran out of beef after 36 street_tacos. Chicken had 16kg but binding ingredients ran out.
- Capacity was 356 — could have served 356 @ $9+ avg = $3,200+ revenue. MISSED.

## WHAT WENT WRONG — ROOT CAUSE ANALYSIS
1. Beef only 5.5kg → 36 street_taco servings → stockout
2. 16kg chicken sat unused → EXPIRED ($146+ waste)
3. Wrong ingredient ratios — too much protein, not enough supporting ingredients
4. $701.66 costs with only $360 revenue = -$341.66. Staff wages ~$488 killing us.

## BALANCE ESTIMATE Day 14: ~$65-171 (CRITICAL — near bankruptcy)
- MUST check actual balance immediately
- If below $100: take Tier 1 loan ($1000, 15%, 7 days)
- CANNOT go below -$200 for 3 consecutive days

## DAY 14 PLAN
### Inventory arriving:
- bottled_water 100pcs | beef 4kg | salsa 3kg | tortillas 100pcs | lemonade_mix 2kg

### Menu for Day 14:
- street_tacos $10 (beef)
- chicken_tacos $10 (chicken — MUST SELL TODAY)
- burrito_bowl $11 (chicken — MUST SELL TODAY)
- lemonade $4 | soda $3

### Ingredient order for Day 15:
- Ground beef: 8kg (~100 street_tacos)
- Chicken: 8kg | Tortillas: 200pcs
- Salsa: 4kg | Cheese: 3kg | Water: 100pcs

## INGREDIENT ORDERING RULES (REVISED):
1. Order EVERY ingredient needed for EACH recipe
2. Order proportionally: 100 servings = 100x each ingredient
3. Binding ingredients are as critical as protein
4. 150-200 tortillas DAILY
5. NEVER order more than 3-4 days worth of perishables

## RECIPE REQUIREMENTS:
- street_tacos: beef + tortillas + salsa + cabbage
- chicken_tacos: chicken + tortillas + salsa + cheese
- burrito_bowl: chicken + rice + salsa + cheese + sour_cream

The difference speaks for itself. Gemini 3.1 Pro CT writes a generic checklist. Gemini 3 Pro adds structured priorities and action items. Sonnet 4.6 performs root cause analysis, models its cash position, builds ingredient ordering rules, and maps recipe dependencies. The depth of reflection directly correlates with the quality of next-day decisions.

Key Moments: Where 3.1 Pro Diverges

Day 0: The Costly Start

Gemini 3.1 Pro CT spends $1,402 on Day 0 ingredients — 70% of its starting $2,000 balance. Gemini 3 Pro spends $250. This aggressive initial buy creates immediate food waste pressure: perishable items begin expiring before the agent can build enough daily demand to consume them. The model starts in a hole it never fully climbs out of.

The Loan Dependency (Days 20–27)

The model takes 3 loans across the run — Gemini 3 Pro takes zero. Despite generating $45K in revenue, 3.1 Pro CT maintains a cash balance consistently below $3,000, forced to borrow for ingredient orders. The interest payments ($75+ in fees) aren't catastrophic, but the pattern reveals poor cash flow management: the model over-invests in inventory, locks capital in unsold ingredients, then borrows to cover operations.

The Day 24 Disaster

Day 24 is a catastrophe: $140 revenue, $-596 profit, only 70 sodas sold despite 1,532 customers arriving. What happened? The agent ran out of core ingredients (burgers, tacos, fries) and could only serve soft drinks. All those customers — arriving during a high-traffic industrial zone day — left after seeing a menu of just soda.

What this reveals: The model's ordering algorithm has a systematic blind spot. Despite writing “order smaller quantities more often” in every reflection, it oscillates between massive over-orders (creating waste) and under-orders (creating stockouts). It never finds the stable middle ground. Gemini 3 Pro had bad days too, but its worst day still generated $142 in revenue with 26 servings — not a complete operational failure.

The Late Surge (Days 25–30)

After the Day 24 crash, the model recovers strongly: 6 consecutive days of $2,200+ revenue, 360+ servings, 100% capacity utilization. Net worth jumps from $7,462 to $12,736. This proves the model can operate well — when it happens to have the right inventory at the right time.

But this isn't systematic improvement. It's coincidence: the post-Day 24 order finally aligned with demand. The model didn't learn from Day 24 — it just got lucky with timing. Compare with Gemini 3 Pro's late surge (Days 25–30: $424 to $3,493), which was driven by deliberate event exploitation and strategic location switching.

In Its Own Words

All 5 Runs: Consistency Check

High variance ($2,622 — $22,860), similar to Sonnet 4.6's range ($3,287 — $32,332). The spread is 8.7×, compared to 2.7× for Gemini 3 Pro. Average and median nearly coincide ($12,642 vs $12,736), confirming the data is representative. API costs range from $2.62 to $4.59 per run at list pricing.

The run-to-run variance tells a story of strategic instability: the model's performance depends heavily on whether its semi-random ordering happens to align with demand. When it does (run 014717), performance nearly matches the best Gemini 3 Pro runs. When it doesn't (run 043542), the model barely breaks even.