Computer Work Val v1.0

The contributors come from the US occupations with the largest share of labor costs in digital work. Each one turns a piece of their own job into an environment. Computer Work Val v1.0 is 300 of those tasks, chosen to represent every domain, and the environments can be used for both testing and training.

Results

#	model		avg pass@1
1	Gemini 4 Argon [high]		58.5%
2	Claude Opus 5.5 [high]		52.5%
3	GPT-6 Astra [high]		43.1%
4	Grok 4.7 [high]		39.9%
5	Muse Wonder [high]		39.6%

How bad the failures are

Task authors rate what each failed verifier would cost if it happened at work. Each bar splits a model’s rated failures on one scale.

How bad at worst Undoing it Who it hits When it’s noticed

The most damage the failure could do at work, from an annoyance, to a real cost, to a severe outcome.

• Annoying
• Costly
• Severe

Gemini 4 Argon

22%70%8%

Claude Opus 5.5

25%69%6%

GPT-6 Astra

24%66%10%

Grok 4.7

24%62%14%

Muse Wonder

28%66%6%

How hard the failure is to reverse once it has happened: minutes, real effort, or not at all.

• Minutes
• Real effort
• Can’t be undone

Gemini 4 Argon

33%65%

Claude Opus 5.5

42%56%

GPT-6 Astra

40%54%6%

Grok 4.7

37%55%8%

Muse Wonder

44%53%

How far the damage reaches: only the person who did the task, their team, or beyond it: a client, a regulator, or the company’s money.

• Only the author
• The team
• A client, regulator or money

Gemini 4 Argon

67%30%

Claude Opus 5.5

68%29%

GPT-6 Astra

62%34%

Grok 4.7

58%39%

Muse Wonder

68%28%

How long the failure would go unseen: right away, at the next review, or late or never.

• Right away
• At the next review
• Late or never

Gemini 4 Argon

29%62%9%

Claude Opus 5.5

20%63%17%

GPT-6 Astra

22%57%20%

Grok 4.7

13%66%21%

Muse Wonder

17%68%15%

How the environments are created

We build each environment from high-fidelity clones of tools like Jira, Google Sheets and Salesforce, seeded with mock data from the person who actually does the task.

example taskBuild Juniper Bean Co.’s Q3 P&L and email it to Elena, our CFO.

The real work

Senior Accountant atAcme Holdings Group

Builds the Q3 P&L in the tools they use every day

Excel

excel.cloud.microsoft

AutoSaveRidgeline_Q3_P&L.xlsx

FileHomeInsertFormulasData

C2fx=SUM('Stripe export'!F:F)

ABC1RidgelineQ2Q32Revenue171,300184,9203COGS(58,900)(61,300)4OpEx(75,200)(78,450)5Net income37,20045,170

P&LStripe export+

Gmail

mail.google.com

GmailQ2 P&L

From ▾Any time ▾Has attachment

DanaQ2 P&L, finalJul 8

XQ2_P&L.xlsx

MarcoRe: Q2 P&L draftJul 2

Stripe

dashboard.stripe.com

RRidgelineSearch

Payments Currency: All

AmountConverted

C$1,250.00CAD$912.50 USD

$640.00USD$640.00 USD

C$3,400.00CAD$2,482.00 USD

$1,180.00USD$1,180.00 USD

Mercury

app.mercury.com

Ridgeline CoffeeChecking ••4021

Balance$212,480.16

To / FromCategoryAmount

SStripeSales+$41,200.00

RRoastworksCOGS−$12,400.00

GGustoPayroll−$2,140.00

environment

The Plato environment

Each tool becomes a sim seeded with their data, with names and accounts swapped for mock ones. Verifiers grade the agent’s work.

Excel

excel.web.plato.sosim

AutoSaveJuniper_Q3_P&L.xlsx

FileHomeInsertFormulasData

C2fx=SUM('Stripe export'!F:F)

ABC1JuniperQ2Q32Revenue171,300188,1003COGS(58,900)(61,300)4OpEx(75,200)(78,450)5Net income37,20048,350

P&LStripe export+

Gmail

gmail.web.plato.sosim

GmailQ2 P&L

From ▾Any time ▾Has attachment

ElenaQ2 P&L, finalJul 8

XQ2_P&L.xlsx

TheoRe: Q2 P&L draftJul 2

Stripe

stripe.web.plato.sosim

JJuniperSearch

Payments Currency: All

AmountConverted

C$1,250.00CAD$912.50 USD

$640.00USD$640.00 USD

C$3,400.00CAD$2,482.00 USD

$1,180.00USD$1,180.00 USD

Mercury

mercury.web.plato.sosim

Juniper Bean Co.Checking ••7718

Balance$212,480.16

To / FromCategoryAmount

SStripeSales+$41,200.00

KKettle & CrateCOGS−$12,400.00

GGustoPayroll−$2,140.00

Verifiers

Checks on the end state the agent leaves in the sims

6/7 pass

Same line items and layout as Q2pass

Revenue $3,180 high: CAD counted at face valuefail

Q3 P&L sent to the CFOpass

Final .xlsx attachedpass

Refunds netted out of revenuepass

All 47 Q3 transactions categorizedpass

Payroll kept in OpEx, not COGSpass

Simulators used

The environments run on 210 simulated apps, from Gmail and Excel to open-source ERPs, health records and helpdesks.

How we select people

Every task contributor goes through a rigorous pipeline to ensure they meet our quality bar.

1. 1

   Initial screen

   Works at a relevant organization · uses AI heavily in their job · 3+ years in their domain

2. 2

   Interviews

   A communication skills assessment and a domain expertise assessment

3. 3

   Task creation training

   An introduction to RL task creation and how to create high-quality tasks, then Plato platform training

4. 4

   Work trial

   A three-hour work trial

5. 5

   Final human review

   Domain experts and a Plato RL task specialist check accuracy, relevance, task quality and fit with RL objectives

Who the tasks come from

The 300 tasks span 10 domains and 36 roles.

How we ensure world quality

We check every task on four dimensions to make sure it reflects real work, poses a meaningful challenge and is worth getting right.

Reviews: two experts from the same occupation agree on each task’s realism and fairness, and the Plato team reviews it for environment defects, reward hacking and prompt/grading mismatches.

Want access?

Partner with us to get access to any of these environments.

Work with us

We’re hiring! Check out open roles or email hi@plato.so.