Acceleration caseWhy sooner
- +1Measured autonomous task horizons have increased rapidly across successive frontier systems.
- +2Coding, tool use, and long-context planning are increasingly designed as one integrated workflow.
Mehan Observatory / Living record 01
The forecast
must survive the evidence.
Six claims about the approaching AI frontier. Each has a probability, a deadline, a disagreement record, and a condition that would prove us wrong.
Enter the probability field →01
The probability field
What we expect.
When we expect it.
Select a point to open its evidence and disagreement record. Probabilities are Mehan Observatory editorial judgments—not scientific consensus.
F-01 / CapabilityRising · review September 2026
62%A week of useful work
By the end of 2028, a frontier AI agent will reliably complete a bounded software or research task that takes a skilled human one working week.
Prior assessment54%→Current62%
Read Disagreement Dossier 01 →In plain language
The important transition is not a smarter answer. It is a machine that can keep working, recover from mistakes, and finish something substantial.
02
The disagreement map
The same evidence.
Two different futures.
A forecast becomes useful when it shows the strongest argument against itself.
Forecast62%
Friction caseWhy later—or never
- −1Performance remains jagged: long runs compound small errors and evaluation awareness can distort results.
- −2A benchmark task is cleaner than real work involving people, changing requirements, and incomplete information.
What would change the assessment?
Independent evaluations showing at least 80% success on unfamiliar, week-scale tasks with no human rescue.
03Where do
Calibrate your judgment
Where do
you stand?
Observatory62%vs.Your estimate62%
You are exactly aligned with the current Observatory assessment.
This comparison stays on your device and is not collected.04Every claim keeps
Open forecast record
Every claim keeps
its revision history.
05Read the record
Evidence desk
Read the record
beneath the judgment.
The selected forecast draws on these primary or independent scientific sources. Links open the underlying evidence; probabilities and interpretations remain the Observatory's own.
International expert writing groupInternational AI Safety Report 2026METRTask-Completion Time HorizonsThe forecast changes when the evidence does
Follow every
revision.
Receive new forecasts, disagreement dossiers, and the evidence that moved our assessment.