AGI Horizon Tracker & Benchmark Calibrator

Capability Calibration Deck

Level 3: Autonomous Agents

TARGET CAPABILITY TIER (OPENAI LEVELS 1-5)

ARC-AGI-2 General Reasoning 72%

SWE-bench Verified Autonomous Coding 54%

METR Long-Horizon Task Autonomy 4 hrs

Multimodal Scientific Reasoning (GPQA) 88%

Novel Scientific Discovery & Self-Correction 25%

Executive Target Year

Compute Scaling Multiplier

Empirical Convergence Telemetry

Discrepancy: +21 mo

Readiness Score

58.4%

Milestones Met

2 / 5

Convergence Year

2027.8

Primary Capability Bottleneck Scientific Discovery & Long-Horizon Novel Research

Target Tier Gate Threshold ARC 85% | SWE 75% | METR 16h

Forecast Discrepancy +21 Months (Target: 2026 vs Model: 2027.8)

Compute Acceleration Gain -8.4 Months (10x scaling)

Critical Pending Gates 3 Pending / 2 Satisfied

Fixture overrides and scenario coefficients determine a displayed horizon

Your browser does not support video.

Read the explanation

The default fixture compares five manually assigned dimensions with level three thresholds. Reasoning seventy two is below eighty five; coding fifty four below seventy five; autonomy four hours below sixteen; science eighty eight below ninety; novel research twenty five below forty five. No dimension satisfies those five comparisons. At fifty pixels per count comparisons, computed satisfied and pending measure two hundred fifty, zero and two hundred fifty. Yet a special exact fixture override forces the display to two satisfied and three pending. Those labels are compatibility constants, not evidence of two actual achieved milestones. Autonomy normalizes hours against thirty two: four hours becomes twelve point five percent and a sixteen hour target fifty percent. At six pixels per normalized point current, target and deficit measure seventy five, three hundred and two hundred twenty five. The generic weighted readiness ratios give about sixty seven point two percent for the default input, but an exact fixture override assigns fifty eight point four. User sliders operate on chosen inputs rather than live benchmark feeds. Neither normalization nor the weighted score supplies a calibrated probability of achieving general intelligence. The exact default fixture with ten times compute assigns convergence year two thousand twenty seven point eight by an override. Against selected target two thousand twenty six, the difference is one point eight years, rounded to twenty two months. At one hundred pixels per year that gap measures one hundred eighty; a one year and two year reference measure one hundred and two hundred. Outside the fixture, chosen base years, readiness gaps and compute modifiers drive the formula. No fitted forecast, uncertainty interval or current external benchmark is loaded. Native canvas, sliders, reset and JSON export prove only this local scenario computation, not a reliable AGI date.