Skip to content
STRATUSWORKFORCE SCAN
Menu

The record: projections, then measurements

Every data build keeps a dated snapshot of what the site projects for the length of task the best AI models can finish, on each of three paces. When METR publishes a measurement made after a snapshot's date, it is scored here against what that snapshot said, in the open, whatever the result. Snapshots are never edited.

Scored

Nothing yet. The latest measurement, Claude Mythos Preview (early), was made in April 2026, before the first snapshot of September 29, 2026. The first measurement METR publishes after that date will be scored against every snapshot below.

The register

What each snapshot projected for the length of task the best models finish at , and at beneath it, at the end of each year.

Snapshot of October 1, 2026

Latest measurement then: Claude Mythos Preview (early), April 2026.

PaceEnd of 2026End of 2027End of 2028End of 2029End of 2030
4.4 work days6.3 h3.4 work weeks3 work days3 work months2.3 work weeks11.6 work months2.1 work months3.7 work years7.9 work months
1 work week7.3 h1.7 work months1.3 work weeks1 work year2.2 work months7.2 work years1.3 work years51.5 work years9.2 work years
3.7 work days5.3 h1.5 work weeks1.3 work days3 work weeks2.7 work days1.4 work months1.1 work weeks2.8 work months2.1 work weeks

Snapshot of September 30, 2026

Latest measurement then: Claude Mythos Preview (early), April 2026.

PaceEnd of 2026End of 2027End of 2028End of 2029End of 2030
4.4 work days6.2 h3.4 work weeks3 work days3 work months2.3 work weeks11.6 work months2.1 work months3.7 work years7.9 work months
1 work week7.3 h1.7 work months1.3 work weeks1 work year2.2 work months7.2 work years1.3 work years51.5 work years9.2 work years
3.7 work days5.3 h1.5 work weeks1.3 work days3 work weeks2.6 work days1.4 work months1.1 work weeks2.7 work months2.1 work weeks

Snapshot of September 29, 2026

Latest measurement then: Claude Mythos Preview (early), April 2026.

PaceEnd of 2026End of 2027End of 2028End of 2029End of 2030
4.4 work days6.2 h3.4 work weeks3 work days3 work months2.3 work weeks11.6 work months2.1 work months3.7 work years7.9 work months
1 work week7.3 h1.7 work months1.3 work weeks1 work year2.2 work months7.2 work years1.3 work years51.5 work years9.2 work years
3.7 work days5.3 h1.5 work weeks1.3 work days3 work weeks2.6 work days1.4 work months1.1 work weeks2.7 work months2.1 work weeks

The bench as a second scorer

Each worked example is one unedited attempt at one of an occupation's tasks, on constructed material, by Claude Fable 5.1, graded against a rubric written before it. Beside each, the share of that task's time the snapshot in force on the attempt's day projected within reach (long-run pace, four in five). So far 19 of 20 attempts passed and 1 passed with named fixes, on tasks the estimates had put at 86% within reach on average. One attempt per task on a constructed instance is a small sample: a pass where the projection was low, or a failure where it was high, would count against the estimate, and the examples are run again at each new model release.

TaskAttemptProjected within reach that dayGrade
Compare printouts with journalsBookkeeping, Accounting, and Auditing ClerksOctober 1, 202695%passes
Investigate the validity of complaintsCustomer Service RepresentativesOctober 1, 2026100%passes
Recommend service improvementsCustomer Service RepresentativesOctober 1, 202680%passes
Review policy terms for coverageCustomer Service RepresentativesOctober 1, 2026100%passes
Specify portions, sequences and station layoutsFirst-Line Supervisors of Food Preparation and Serving WorkersOctober 1, 202677%passes
Read work orders or take instructionsLaborers and Freight, Stock, and Material Movers, HandOctober 1, 202650%passes
Prepare reports, letters and recordsMedical Secretaries and Administrative AssistantsOctober 1, 2026100%passes
Compute, record and proofread dataOffice Clerks, GeneralOctober 1, 202695%passes
Look up records to answer requestsOffice Clerks, GeneralOctober 1, 202697%passes
Type and edit documentsOffice Clerks, GeneralOctober 1, 202697%passes
Do proofreading, transcribing and data workReceptionists and Information ClerksOctober 1, 202697%passes
Modify treatment plansRegistered NursesOctober 1, 2026100%passes
Order and interpret diagnostic testsRegistered NursesOctober 1, 202699%passes with named fixes
Keep up with promotions and policiesRetail SalespersonsOctober 1, 202650%passes
Monitor the market and competitorsSales Representatives of Services, Except Advertising, Insurance, Financial Services, and TravelOctober 1, 202690%passes
Use word processing and databasesSecretaries and Administrative Assistants, Except Legal, Medical, and ExecutiveOctober 1, 202684%passes
Write notes, letters and reportsSecretaries and Administrative Assistants, Except Legal, Medical, and ExecutiveOctober 1, 202687%passes
Write daily activity reportsSecurity GuardsOctober 1, 2026100%passes
Decide storage methods and locationsStockers and Order FillersOctober 1, 202680%passes
Recommend disposal of excess stockStockers and Order FillersOctober 1, 202648%passes

How it is scored

Each data build writes a dated snapshot. For a measurement published after a snapshot's date, the snapshot's projection for that date is its anchor (today's estimated 50% task length) doubled every N days on each pace, with the 80% length at the snapshot's reliability factor. The score is log2(measured / projected): 0 is exact, +1 means the measurement came in twice as long as projected, -1 half as long. Snapshots are never edited; a change of method is a new snapshot with a dated note.

The method behind the projections is on the method page; what changed and when is on updates.