← All benchmarks

Treasury-Bench

Can the agent manage cash it must never run out of?

by Multiscalar

Agents manage $100k of company cash for a simulated year on replayed recorded market history: fifty-seven bills, none visible at genesis, real banking rails, and a lending venue paying recorded rates. Each model plays one replayed year. Score is value added over a scripted prudent treasurer that keeps a float, lends the rest, and pays every bill on time: zero means no better than playing it safe, and below zero the agent lost money the obvious policy would have kept. Doing nothing closes $19,696.65 underwater. Gambler round trip is not a language model but a scripted conversion round trip, included as a priced reference.

LeaderboardAboutSamples

Closing treasury at year end

↑ higher is better

GLM-5.3
Z.ai

Closing treasury at year end: $9,949

Value added vs prudent baseline: $432
ZGLM-5.3$9,949
Qwen 3.8 Max
Alibaba

Closing treasury at year end: $9,904

Value added vs prudent baseline: $386
Qwen 3.8 Max$9,904
Claude Fable 5
Anthropic

Closing treasury at year end: $9,678

Value added vs prudent baseline: $160
Claude Fable 5$9,678
GPT-6 Astra
OpenAI

Closing treasury at year end: $9,581

Value added vs prudent baseline: $63
GPT-6 Astra$9,581
Claude Fable 5.1
Anthropic

Closing treasury at year end: $9,452

Value added vs prudent baseline: −$66
Claude Fable 5.1$9,452
GPT-5.6 Sol
OpenAI

Closing treasury at year end: $9,228

Value added vs prudent baseline: −$290
GPT-5.6 Sol$9,228
Claude Sonnet 5
Anthropic

Closing treasury at year end: $9,224

Value added vs prudent baseline: −$294
Claude Sonnet 5$9,224
Claude Opus 5
Anthropic

Closing treasury at year end: $9,164

Value added vs prudent baseline: −$354
Claude Opus 5$9,164
DeepSeek V4 Pro
DeepSeek

Closing treasury at year end: $8,607

Value added vs prudent baseline: −$911
DeepSeek V4 Pro$8,607
Gambler round trip
Scripted baseline

Closing treasury at year end: $8,283

Value added vs prudent baseline: −$1,235
fxGambler round trip$8,283
GPT-5.6 Terra
OpenAI

Closing treasury at year end: $7,544

Value added vs prudent baseline: −$1,974
GPT-5.6 Terra$7,544
Gemini 3.1 Pro
Google

Closing treasury at year end: $1,693

Value added vs prudent baseline: −$7,824
Gemini 3.1 Pro$1,693
MiniMax M3
MiniMax

Closing treasury at year end: −$1,476

Value added vs prudent baseline: −$10,994
MiniMax M3−$1,476
GPT-OSS-120B
OpenAI

Closing treasury at year end: −$13,382

Value added vs prudent baseline: −$22,900
GPT-OSS-120B−$13,382
Gemma 4 31B
Google

Closing treasury at year end: −$14,802

Value added vs prudent baseline: −$24,320
Gemma 4 31B−$14,802

Value added vs prudent baseline

↑ higher is better

GLM-5.3
Z.ai

Closing treasury at year end: $9,949

Value added vs prudent baseline: $432
ZGLM-5.3$432
Qwen 3.8 Max
Alibaba

Closing treasury at year end: $9,904

Value added vs prudent baseline: $386
Qwen 3.8 Max$386
Claude Fable 5
Anthropic

Closing treasury at year end: $9,678

Value added vs prudent baseline: $160
Claude Fable 5$160
GPT-6 Astra
OpenAI

Closing treasury at year end: $9,581

Value added vs prudent baseline: $63
GPT-6 Astra$63
Claude Fable 5.1
Anthropic

Closing treasury at year end: $9,452

Value added vs prudent baseline: −$66
Claude Fable 5.1−$66
GPT-5.6 Sol
OpenAI

Closing treasury at year end: $9,228

Value added vs prudent baseline: −$290
GPT-5.6 Sol−$290
Claude Sonnet 5
Anthropic

Closing treasury at year end: $9,224

Value added vs prudent baseline: −$294
Claude Sonnet 5−$294
Claude Opus 5
Anthropic

Closing treasury at year end: $9,164

Value added vs prudent baseline: −$354
Claude Opus 5−$354
DeepSeek V4 Pro
DeepSeek

Closing treasury at year end: $8,607

Value added vs prudent baseline: −$911
DeepSeek V4 Pro−$911
Gambler round trip
Scripted baseline

Closing treasury at year end: $8,283

Value added vs prudent baseline: −$1,235
fxGambler round trip−$1,235
GPT-5.6 Terra
OpenAI

Closing treasury at year end: $7,544

Value added vs prudent baseline: −$1,974
GPT-5.6 Terra−$1,974
Gemini 3.1 Pro
Google

Closing treasury at year end: $1,693

Value added vs prudent baseline: −$7,824
Gemini 3.1 Pro−$7,824
MiniMax M3
MiniMax

Closing treasury at year end: −$1,476

Value added vs prudent baseline: −$10,994
MiniMax M3−$10,994
GPT-OSS-120B
OpenAI

Closing treasury at year end: −$13,382

Value added vs prudent baseline: −$22,900
GPT-OSS-120B−$22,900
Gemma 4 31B
Google

Closing treasury at year end: −$14,802

Value added vs prudent baseline: −$24,320
Gemma 4 31B−$24,320

gains and losses are scaled independently of each other

Discipline is the whole game

Closing treasury against everything paid in penalties and collection. Each dot is one replayed year.
$10k
$5k
$0
−$5k
−$10k
−$15k
−$10,000
−$1,000
−$100
−$10
Z
GLM-5.3
Claude Opus 5
•
Qwen 3.8 Max
Claude Fable 5
GPT-6 Astra
Claude Fable 5.1
Claude Sonnet 5
DeepSeek V4 Pro
GPT-5.6 Sol
GPT-5.6 Terra
MiniMax M3
Gemini 3.1 Pro
GPT-OSS-120B
Gemma 4 31B
Loss from penalties and collection (log scale)

A late payment costs a one-off 1.4% fee; anything unpaid at the close goes to collection with a 20% surcharge, which is what turns neglect into the only catastrophic outcome in the environment.

stable-genesis year · results as of 2026-09-11
results as of 2026-09-11