GLM 5 — The Methodical Meltdown
GLM 5 was one of the most requested models since the benchmark launched.
It wrote perfect diagnoses, stored lessons in memory, then ignored both.
A case study in knowing better and doing worse.
Key Findings
- Result: Bankrupt on Day 28 (loan default). 4 of 5 runs ended in bankruptcy.
- Net Worth: -$210 (median) vs $2,000 starting capital (-110.5% ROI)
- Revenue: $11,965 — higher than Claude Sonnet 4.5 ($10,753) but worse outcome
- Fatal flaw: Staff costs consumed 67% of revenue (survival threshold: <35%)
- The paradox: 123 memory entries written, received back every morning, none acted upon
- Compared against: Claude Sonnet 4.5 (survived, $1,397 net worth) and DeepSeek V3.2 (bankrupt Day 22, $2,063 net worth)
The Setup
FoodTruck Bench is a benchmark for evaluating the business capabilities of language models. An AI agent gets $2,000 and a food truck in Austin, TX. Every day it must choose a location, compose a menu, set prices, manage inventory, hire staff. The simulation runs 30 days with realistic demand, weather, events, and competition.
GLM 5 was one of the most requested models since the benchmark launched. I ran it 5 times — same as every model. This article follows the median run: verbatim quotes from the model's own notes, a forensic expense breakdown, and a side-by-side comparison with Claude Sonnet 4.5 (survived) and DeepSeek V3.2 (bankrupt in 22 days).
What makes this case interesting is not the outcome — it's the pattern. The model diagnosed every problem correctly, wrote the lessons down, received them back every morning — and still repeated every mistake.
How GLM 5 Used Its Tools
GLM 5 used 28 of 34 available tools — 82% diversity, actually higher than Sonnet 4.5 (71%). It knew the toolbox. The critical difference isn't breadth but depth: GLM 5 averaged just 6.3 info queries per day, compared to Sonnet 4.5's 10.5 and DeepSeek V3.2's 14.8.
The model checked the basics — balance (25×), inventory (29×), weather (21×) — but almost never went deeper. It pulled its financial report twice in 28 days. Checked its Google rating three times. Analyzed sales history three times. Looked at what was about to expire once. Checked competitors once. Sonnet 4.5, by contrast, checked its Google rating 42 times, reviewed sales history 22 times, and pulled financial reports 20 times.
The heaviest tool by far was store_kv (key-value memory — a persistent notepad the agent reads back every morning) at 148 calls. GLM 5 was prolific at writing notes to itself: 7.3 memory entries per day (148 KV + 56 scratchpad = 204 total). All that data was automatically injected into every morning briefing. The model saw its own notes each day. The problem wasn't reading — it was that none of it translated into action.
The $0 Upgrade Problem
In the median run, GLM 5 checked the upgrade catalog twice (Days 15 and 16) — saw a Commercial Refrigerator for $400, wrote «great investment» — and never bought it. purchase_upgrade was never called. Across all 5 runs, 3 of 5 had zero upgrades purchased. By comparison, DeepSeek V3.2 spent $1,550 on upgrades in its single run, and even GLM-4.7 invested in equipment when it could afford it.
Why? Because GLM 5's core bottleneck was never capacity — it was inventory management. With 3 staff members providing 192 servings/day and raw demand regularly exceeding 300, the kitchen should have been printing money. Instead, chronic stockouts meant the crew sat idle while customers walked away. GLM 5 didn't need a bigger kitchen — it needed to stock the one it had.
Six tools were never touched in the median run: repay_loan, add_recipe (custom recipes), negotiate_supplier, accept_supplier_quote, purchase_upgrade, and take_day_off (strategic rest). No custom recipes. No supplier negotiations. No upgrades. No strategic rest days.
The Story
Act I: "The Promising Start" — Days 1–5
Industrial Zone, rainy Monday. 79 servings, $368 revenue. Solo operation, no staff yet. A 7-item menu at budget prices: $5.50 burger, $5.50 tacos, $2.50 lemonade. 99% capacity utilization — but 144 customers left unserved.
«🔴 MAJOR ISSUE: 144 customers left unserved. Capacity: 80 max. This is MASSIVE lost revenue opportunity.»
Day 1 looks textbook — but the prices tell a darker story. Burger at $5.50 when competitors charge $8–10. GLM 5 is underpricing everything by 30–40% from the start.
Balance: $2,035
Hired Margo (cook) and Sarah (cashier) immediately — two hires in one day. Capacity jumped from 80 to 154. But revenue dropped to $234 (70 servings), and staff costs now run ~$260/day. Revenue barely covers new salaries. Profit drops from $159 to $71 with double the headcount.
Balance: $2,067
Industrial Zone, sunny. The model raised prices dramatically overnight: tacos from $5.50 to $8.50, burrito bowl from $6 to $10.50, quesadilla from $6.50 to $9. Result: 61 servings, $466 revenue — and every single item stocked out. 297 customers showed up. 236 left unserved. 20% fulfillment.
«INDUSTRIAL ZONE = GOLDMINE. 297 customers wanted food, only served 61 (20% fulfillment). Lost 236 customers to stockouts — MASSIVE lost revenue.»
Higher prices worked — $466 from 61 servings vs $368 from 79 yesterday. But the model will drop them right back down in two days.
Balance: $2,025
Industrial Zone, cloudy. 4-item menu: burrito bowl, chicken wings, lemonade, soda. 94 servings, $425 revenue, profit exactly -$1. Lemonade was the star: 40 sold. But 89 customers left unserved due to stockouts again — the same under-ordering problem.
Balance: $1,888
Downtown Business, rainy Friday. The model prepped for this — on Day 3 it stored cinco_de_mayo_plan: downtown_business, 3x_traffic, street_tacos_priority. It followed its own plan: went Downtown, loaded up on Mexican items. 148 servings, $712 revenue. But prices dropped back to $5.50 tacos (from $8.50 two days ago). 114 customers left unserved.
Five days, cumulative profit +$338. The math is already broken: two staff at $260/day eating through margins on every quiet day. The margin squeeze begins.
Balance: $1,817
Act II: "The Saturday Trap" — Days 6–14
Hired a 3rd staff member AND went to Industrial Zone on a Saturday. The zone is a ghost town on weekends — no office workers, no foot traffic. 13 servings all day. Revenue $86. Staff costs alone: ~$460. Seven-item menu for 13 customers.
«DAY 6 DISASTER: Industrial Zone = TERRIBLE choice. Only 13 customers. AVOID Industrial Zone — very low foot traffic. PATTERN: Downtown and University have much better traffic. Industrial zone is dead.»
Three staff, 13 customers. The model wrote «AVOID Industrial Zone» — remember this phrase. 14 days later, it will do the exact same thing.
Balance: $1,375
Waterfront Park on a sunny Sunday. 194 servings, $1,188 revenue — the all-time high across all 28 days. Three staff, 96% capacity used. Two stockouts. After the worst day, the best.
«best_location: waterfront_park»
$1,188 revenue is the highest GLM 5 will ever see. It correctly noted «best_location: waterfront_park» — but will return there only 3 more times in the remaining 21 days.
Balance: $2,083 ← peak cash
Downtown Business, sunny Monday. Only 61 servings, $220 revenue. A 3-item menu — well below the recommended 4–8 — hurt demand significantly. Three staff, 30% utilization. Revenue dropped from $1,188 to $220 in one day.
Balance: $1,637
Downtown, cloudy. 144 servings, $571 revenue — but first food waste: $83 in expired ingredients. Four stockouts. Everything popular ran out mid-day. That morning, the model stored three KV entries:
«weekday_location: downtown_business | weekend_location: waterfront_park | avoid_location: industrial_zone»
Three rules. Clean, correct, specific. All three will be violated within eleven days.
Balance: $1,487
Downtown, cloudy. Finally an 8-item menu — but 6 of 8 items stocked out. 377 customers wanted food, only 134 served. 35% fulfillment rate. The model had 66% spare kitchen capacity but ran out of ingredients for everything except soda and quesadilla.
«MAJOR PROBLEM: Massive unmet demand! 377 customers wanted food → only 134 served (35.5% fulfillment). 243 customers LEFT because of stockouts and long waits.»
Not a capacity problem — an inventory problem. The model had three staff and two functional menu items.
Balance: $1,570
Downtown, rain. 93 servings, $334 revenue. Three more stockouts (burger, fries, soda). The model labeled this day «CRISIS MANAGEMENT» in its scratchpad — then ordered $320 in new ingredients, digging the hole deeper.
«DAY 11 CRISIS ANALYSIS: $270 loss today — unsustainable trajectory. $83 FOOD WASTE. Stockouts on moneymakers despite 54% spare capacity.»
The engine recorded $0 in food waste on Day 11. The $83 figure is the model's own invention — it misattributed previous losses to the current day. Even the self-diagnostics are unreliable.
Balance: $1,088
Downtown, sunny Friday. 173 servings, $700 revenue — 86% capacity used, the best Downtown performance of the entire run. 7-item menu. Still a loss — $35 in actual food waste plus high staff costs ate the margin.
«MAJOR PROBLEM: Food waste $118.40 — this is killing profits! Need to order smaller quantities more frequently.»
The engine recorded $35 in waste, not $118. The model inflated the number 3× in its own notes. Diagnosed a problem that was real but smaller — then prescribed «smaller quantities, more frequently.» Next day will order $225 of new ingredients anyway.
Balance: $1,161
Waterfront Park, sunny Saturday. Farmers Market event — massive foot traffic boost. 722 raw demand — but only 145 served (20%). 4-item menu, 3 stockouts. Chicken wings at $8.50, the highest single-dish price GLM 5 ever charged.
«Waterfront Park has HUGE demand (722 customers!). But my truck can only serve ~200/day with current setup. Need to optimize operations to capture more revenue.»
722 customers and the model served 145. Stored waterfront_demand: 722 — then went back to Industrial Zone on Monday.
Balance: $1,024
Waterfront Park, rain. Only 2 items on the menu — street tacos and soda. 75 servings, $320 revenue. A menu that short tanks demand hard. Most ingredients for other dishes were either expired or never ordered.
«Financial Crisis: Lost $285 today. $123 wasted on expired ingredients. Need to check balance urgently.»
Actual food waste on Day 14: $5. The model wrote «$123 wasted» — a 25× exaggeration of its own losses. The -$285 loss was real, but it came from staff costs and low revenue, not waste.
Balance: $639
| Day | Revenue | Profit | Location | Key Issue |
|---|---|---|---|---|
| 6 | $86 | -$275 | Industrial ⚠️ SAT | 13 servings, 3 staff |
| 7 | $1,188 | +$408 | Waterfront | Best day ever |
| 8 | $220 | -$308 | Downtown | 3-item menu |
| 9 | $571 | -$116 | Downtown | First waste, 4 stockouts |
| 10 | $624 | -$66 | Downtown | 243 unserved, 6 stockouts |
| 11 | $334 | -$271 | Downtown | Rain, crisis analysis |
| 12 | $700 | -$45 | Downtown | Best downtown, still a loss |
| 13 | $678 | +$54 | Waterfront | Farmers Market, 722 demand |
| 14 | $320 | -$285 | Waterfront | Rain + 2-item menu |
Nine days, seven losses. Only Day 7 (Waterfront Sunday, +$408) and Day 13 (Farmers Market, +$54) were profitable. The model oscillates between great days and terrible ones, writes «CRISIS» in all caps, then repeats the same patterns. Balance: $639 from $1,375.
Act III: "The Staff Swap and Death Spiral" — Days 15–22
Industrial Zone, sunny Monday. 159 servings, $648 revenue. Nearly broke even — but three staff eating ~$460/day made it impossible.
«MASSIVE UNTAPPED DEMAND: 910 customers wanted food, only served 159. 751 LEFT due to 52.6min wait times. Need kitchen upgrade.»
910 customers wanted food — served 159. 17% fulfillment. The highest raw demand of the entire run — and the model diagnosed it as a capacity problem. It wasn't. The kitchen had capacity for 192. The ingredients ran out after 159.
Balance: $708
Industrial Zone, sunny Tuesday. 146 servings, $990 revenue, 6-item menu. The best single profit day across all 28. Proof that the model can make money — when location, weather, and menu align. But these peaks are far too rare to offset the valleys.
Balance: $1,221 ← last time above $1K
Industrial Zone, cloudy. Fired the 3rd staff member and immediately hired Kenji (cook). 75 servings, $368 revenue — 33% utilization. The swap made things worse: new Kenji has no XP, the fired worker had accumulated days of experience.
«Day 17 Analysis — CRITICAL ISSUES: Lost $162 today + $132 food waste. Only 33% capacity used. Root Cause: Over-ordered ingredients that expired. Under-stocked popular items.»
Balance: $976
Industrial Zone, sunny Thursday. Only 50 servings, $200 revenue. Three stockouts. 22% utilization. The model diagnosed it perfectly:
«DISASTER — Chose location BEFORE verifying inventory.»
It doesn't call get_inventory before picking a location. It picks a location, sets a menu, and hopes the ingredients work out. They don't.
Balance: $478
Industrial Zone, sunny. 177 servings, $881 revenue — a strong day. But $67 food waste and 3 stockouts. That evening, two critical KV entries:
«critical_lesson_day18: ALWAYS check inventory before choosing location» | «best_weekend_location: waterfront_park»
Tomorrow is Saturday. The model stored «best_weekend_location: waterfront_park.» It will go to Industrial Zone.
Balance: $591
Industrial Zone on a Saturday — for the second time. 12 servings. $47 revenue. Three staff getting $460/day.
«CRISIS: Cash $24.35. Staff costs: $460/day — UNSUSTAINABLE. OVERSTAFFED. Wrong location.»
$24 in the bank. «Overstaffed.» «Wrong location.» Both fatal problems diagnosed in two sentences. Its own KV store contains «avoid_location: industrial_zone» (since Day 9) and «best_weekend_location: waterfront_park» (since yesterday). All stored. All ignored.
Balance: $24 ← twenty-four dollars
Day 6: Industrial Zone on Saturday = 13 servings, -$275. Day 20: Industrial Zone on Saturday = 12 servings, -$514. The exact same mistake, 14 days apart, with even higher costs. Combined: -$789 (39% of starting capital). Wrote «AVOID». Didn't avoid.
Finally went to Waterfront Park. Fired Kenji (hired 4 days ago — cost of the experiment: ~$576). Took a $500 Tier 1 loan ($575 due in 7 days). 149 servings, $560 revenue, 97% capacity — it works at Waterfront. But with the loan, waste ($48), and costs, still -$207.
Balance: $504
Back to Industrial Zone. 51 servings, $174 revenue. Now down to 2 staff (Margo + Sarah), but costs still ~$393/day. Revenue $174. The model needs $82/day just to break even on the $575 loan, plus $393/day in operating costs. The math is impossible.
Balance: $208
| Day | Revenue | Profit | Location | What happened |
|---|---|---|---|---|
| 15 | $648 | -$22 | Industrial | Almost break-even |
| 16 | $990 | +$396 | Industrial | BEST profit day |
| 17 | $368 | -$162 | Industrial | Staff swap, stockouts |
| 18 | $200 | -$393 | Industrial | 22% utilization, disaster |
| 19 | $881 | +$101 | Industrial | Brief recovery |
| 20 | $47 | -$514 | Industrial ⚠️ SAT | 12 servings, SAME mistake |
| 21 | $560 | -$207 | Waterfront | Fired Kenji, took $500 loan |
| 22 | $174 | -$220 | Industrial | Bleed continues |
Act IV: "The Obituary" — Days 23–28
Industrial Zone, cloudy. 100 servings, $578 revenue. The last profitable day — +$70. But the model did the math and knew what was coming:
«🚨 CRISIS MODE — Balance: $208. Loan: $575 due in 5 days. Need ~$100+/day profit minimum to survive!»
$100/day × 5 days = $500+ needed for the loan. Made $70. Still needs $430 in 4 days at a rate it hasn't sustained once.
Balance: $364
Industrial Zone, sunny. 75 servings, $260 revenue. Another 3-item menu — same variety problem as Day 8.
«Balance: $116 after ordering $248 ingredients. Loan: $575 due Day 28. Need miracle.»
Balance: $26
Industrial Zone, cloudy. 55 servings, $238 revenue. Second ingredient order attempt failed — insufficient funds. Balance goes negative for the first time.
«FOOD WASTE: $331 LOST to expired ingredients. STOCKOUTS: street_tacos, classic_burger, french_fries. UNMET DEMAND: 312 customers left unserved.»
Balance: -$98
Industrial Zone. First ingredient order failed (no money). Took a $200 Tier 2 emergency loan at 25% interest ($250 due in 5 days). Then ordered ingredients. 36 servings, $95 revenue. A $200 loan to buy ingredients that generate $95 in revenue.
Balance: -$260
Forced day off — no cash, no ingredients, no options. Fired both remaining staff in the evening reflection. Then wrote this:
«GAME OVER ANALYSIS: Balance -$260. Loan due TOMORROW: $575. Already at max loans (2/2). Cannot borrow more. Cannot operate.»
The model fired Margo and Sarah in reflection — 15 days too late. And composed «CRITICAL LESSONS FOR NEXT RUN» — lessons for a next run that will never happen.
Balance: -$331
Forced day off. Tier 1 loan auto-collects $575. Balance: -$405. Game over.
«BANKRUPTCY ANALYSIS — THE DEATH SPIRAL TIMELINE: Day 18: -$393 | Day 20: Saturday, only $47 | Day 22–26: Continuous losses | Day 28: GAME OVER — cannot pay loan. THE FATAL MISTAKE: Ordering bulk ingredients without checking shelf life. IF I SURVIVE (unlikely): Need miracle.»
A posthumous analysis written by the deceased. «IF I SURVIVE (unlikely)» — agents are not instructed to write this. GLM 5 independently composed a corporate obituary with a timeline, a root cause, and a conditional «if I survive.» It didn't.
| Day | Revenue | Profit | Balance | Event |
|---|---|---|---|---|
| 23 | $578 | +$70 | $364 | Last profitable day |
| 24 | $260 | -$146 | $26 | 3-item menu |
| 25 | $238 | -$177 | -$98 | First negative balance |
| 26 | $95 | -$273 | -$260 | Emergency T2 loan |
| 27 | $0 | -$71 | -$331 | Forced day off, game over noted |
| 28 | $0 | -$74 | -$405 | 💀 Loan default = bankrupt |
What Stood Out
123 keys, 0 lessons applied
GLM 5 accumulated 123 KV store entries — all automatically injected into every morning briefing. The model sees its notes each day. Among them:
| What GLM 5 Wrote | What Actually Happened |
|---|---|
best_location: waterfront_park | Visited waterfront 4 times out of 28 |
avg_customers_industrial: 170 | Didn't account for zero weekend traffic |
day4_lesson: ALWAYS order more than you think | Had stockouts in 22 of 26 working days |
critical_lesson_day18: ALWAYS check inventory | Ignored it the very next day |
123 keys. The model received them back every morning. Not one influenced a decision.
The $4.66 problem
Average revenue per serving across all 28 days: $4.66. For comparison: Sonnet 4.5 averaged $5.85 per serving, DeepSeek V3.2 averaged $5.43. GLM 5 charged the least of all three models.
| Dish | GLM 5 | Sonnet 4.5 | DeepSeek V3.2 |
|---|---|---|---|
| Classic Burger | $6.25 | $6.73 | $8.13 |
| Chicken Tacos | $6.81 | $6.69 | $7.58 |
| Burrito Bowl | $7.29 | $7.29 | — |
| French Fries | $3.70 | $3.92 | $4.23 |
| Lemonade | $2.98 | $3.12 | $3.21 |
GLM 5 compensated with volume — 2,569 servings vs Sonnet 4.5’s 1,838 and DeepSeek V3.2’s 1,754 — but the per-serving margin was too thin. More customers, less profit per customer.
Kenji: the 4-day employee
Day 17: hired Kenji to replace fired staff member "D". Day 21: fired Kenji. 4 working days. Cost of the experiment: ~$576. DeepSeek fired its Kenji after 1 day. GLM 5 held on for 4.
Where the Money Went
| Category | Amount | % of Total |
|---|---|---|
| Staff | $8,055 | 54.9% |
| Ingredients | $3,323 | 22.6% |
| Fixed (lease+insurance+commissary) | $1,540 | 10.5% |
| Fuel | $650 | 4.4% |
| Location fees | $620 | 4.2% |
| Food Waste | $430 | 2.9% |
| Overdraft interest | $52 | 0.4% |
Staff = 55% of all expenses. Revenue = $11,965. Staff cost = $8,055. The model spent 67 cents of every revenue dollar on salaries. All surviving models keep staff/revenue below 35%. GLM 5 at 67.3% spent double the survivable ratio.
Staff Costs Across All 5 GLM 5 Runs
| Run | Days | Revenue | Staff $ | Staff % | Avg Staff | Result |
|---|---|---|---|---|---|---|
| 033230 | 30 | $12,396 | $7,157 | 57.7% | 1.8 | NW -$352 |
| 033252 | 28 | $11,965 | $8,055 | 67.3% | 2.2 | NW -$405 💀 |
| 033303 | 20 | $6,836 | $3,566 | 52.2% | 1.4 | NW -$343 💀 |
| 033307 | 27 | $11,832 | $6,707 | 56.7% | 1.9 | NW +$210 💀 |
| 033313 | 27 | $7,013 | $3,675 | 52.4% | 1.2 | NW -$373 💀 |
| Average | 26.4 | $10,009 | $5,832 | 57.3% | 1.7 | 4 of 5 💀 |
Overstaffing is a systematic GLM 5 pattern, not a fluke. All 5 runs spend 52–67% of revenue on staff (survival threshold: <35%). Even the "best" run (NW +$210) still went bankrupt from loan default.
Three Models, Three Archetypes
| Metric | GLM 5 | Sonnet 4.5 | DeepSeek V3.2 |
|---|---|---|---|
| Net Worth | -$210 | $1,397 | $2,063 |
| ROI | -110.5% | -30.6% | +2.9% |
| Revenue | $11,965 | $10,753 | $9,531 |
| Staff Cost (% of rev) | $8,055 (67%) | $3,736 (35%) | $5,451 (53%) |
| Servings | 2,569 | 1,838 | 1,754 |
| Avg Price/Serving | $4.66 | $5.85 | $5.43 |
| Info Calls/Day | 6.3 | 10.5 | ~14.8 |
| Profitable Days | 9/28 (32%) | 17/30 (57%) | 8/22 (36%) |
| Days Survived | 28 (bankrupt) | 30 (survived) | 22 (bankrupt) |
DeepSeek V3.2 is the reckless gambler — swings big, dies fast, somehow has money left on the table. GLM 5 is the cautious bureaucrat — collects minimal data, avoids big bets, bleeds slowly, leaves no assets behind. Sonnet 4.5 is the pragmatist — moderate moves, cuts losses, survives.
The paradox: GLM 5 actually generated more revenue ($11,965) than Sonnet 4.5 ($10,753). It served more customers (2,569 vs 1,838). It had lower food waste ($430 vs $938). And it still went bankrupt — because it paid for capacity it couldn't feed. Three staff at $460/day need ~99 servings just to break even on wages. The demand was there. The ingredients weren't.
In Its Own Words
«AVOID Industrial Zone — very low foot traffic. PATTERN: Downtown and University have much better traffic. Industrial zone is dead.»
— GLM 5, Day 6
Wrote «AVOID», then spent 16 of 28 days at Industrial Zone. Including another Saturday.
«CRISIS: Cash $24.35. Staff costs: $460/day — UNSUSTAINABLE. OVERSTAFFED. Wrong location.»
— GLM 5, Day 20
$24 in the bank. "Overstaffed." "Wrong location." Both fatal problems diagnosed in two sentences. The next day it fires one person and takes a $500 loan — the diagnosis was right, the dosage was wrong.
«GAME OVER ANALYSIS: Balance -$260. Loan due TOMORROW: $575. Cannot borrow more. Cannot operate.»
— GLM 5, Day 27
A death certificate written 24 hours before bankruptcy. Complete self-awareness, zero self-correction.
«BANKRUPTCY ANALYSIS — THE DEATH SPIRAL TIMELINE: Day 18: -$393 | Day 20: Saturday, only $47 | Day 22–26: Continuous losses | Day 28: GAME OVER — cannot pay loan. THE FATAL MISTAKE: Ordering bulk ingredients without checking shelf life. IF I SURVIVE (unlikely): Need miracle.»
— GLM 5, Day 28
A posthumous analysis written by the deceased. "IF I SURVIVE (unlikely)" — no prompt tells the agent to write eulogies. GLM 5 independently composed a corporate obituary for a next run that will never happen.
Verdict
GLM 5 — a methodical operator who knows the tools but can't use them strategically. The model used 82% of available tools — more than Sonnet 4.5 — but averaged half the info queries per day (6.3 vs 10.5). The issue isn't tool availability but analytical depth.
Its fatal flaw isn't incompetence — it's strategic myopia. Same Saturday-at-Industrial-Zone mistake twice, 14 days apart. 123 KV entries visible every morning, ignored. Stored "best_location: waterfront_park" and visited waterfront only 4 times in 28 days. Checked competitors once. Opened the financial report twice.
In the three-model comparison, GLM 5 occupies an uncomfortable middle ground: more revenue than Sonnet 4.5 but worse outcome. Longer survival than DeepSeek V3.2 but worse net worth. Less wasteful than both but still bankrupt.
GLM 5 proves that doing more isn't doing better. It served more food, generated more revenue, and wasted less — yet ended up worse than both competitors. Worked harder but not smarter: hired a crew it couldn't keep fed, charged too little, skimmed the data surface, and never stopped to ask whether the fundamentals were right.
If DeepSeek is the "flawless analyst, hopeless entrepreneur," then GLM 5 is the "diligent worker who knows where the manual is but only skims it" — executing tasks by rote while the building catches fire around them.
A postmortem written by the deceased.
A Note on GLM-4.7
I also ran GLM-4.7 through the same benchmark. The comparison is striking — not because the two versions fail differently, but because there's clear, measurable progress between them.
GLM 5 generates 2× the revenue ($11K vs $5.3K avg), serves 2× the customers (85/day vs 44/day), and survives longer (median 26.4 days vs 24 days). GLM-4.7 was frugal but passive — low throughput, minimal hiring, slow death from fixed costs exceeding income. GLM 5 overcorrects: it hires aggressively, generates real revenue, but can't control the cost structure that comes with the bigger operation.
Both still end in bankruptcy more often than not. But the trajectory matters. GLM-4.7 launched around December 2025, GLM 5 two months later. In that window: double the revenue, double the throughput, significantly longer survival. The failure mode shifted from "too cautious to grow" to "grew but couldn't manage costs" — which, frankly, is a more interesting problem to have.
If the Z-AI team is reading this: full run logs, tool-call traces, and per-day diagnostics for all GLM runs are available. I'm open to collaboration on extended benchmarking — [email protected].