Open research
The clearest measure of AI progress is not a leaderboard score — it is the length of work an AI can finish on its own. That horizon now reaches ~16 hours of expert work at 50% reliability, and it has been doubling every ~4-5 months. This page tracks that exponential against the human baseline, and the feedback loop — AI accelerating AI R&D — that drives it. Every number is sourced. Extrapolations are not forecasts.
The capability exponential
The length of human-expert work a frontier model finishes at 50% reliability (METR-style). Points are measured results; the solid fit is a least-squares exponential; the dashed line extrapolates the recent doubling toward week-, month-, and year-long autonomous tasks.
Now (2026-05)
~16 hours
at 50% reliability · doubling ~4-5 months
1 work-day
8hReached
as of 2026-05
1 work-week
40h2026
projected · band 2026–2027
1 month
160h2027
projected · band 2027–2034
1 year
1,920h2029
projected · band 2028–2037
Fit: least-squares regression of log2(task minutes) on year gives a doubling time of ~5.2 months across the measured era (7-month historical, ~4.5-month recent). Extrapolation is not a guarantee — stricter 80%-reliable horizons lag, and a compute-scaling slowdown sets the high end of each band.
Crossing the human line
AI has surpassed expert-human baselines on 4 of 6 tracked benchmarks, from broad knowledge (MMLU) to PhD-level science (GPQA) and competition math. Real computer-use agents are now at human parity.
ARC-AGI-2 — novel abstraction that resists memorization — is the live frontier, crossing the human panel only at high compute. Research-grade mathematics (FrontierMath) is the clearest remaining headroom.
Broad academic knowledge
SurpassedSaturated near and above the human-expert reference; retired as a frontier signal.
Sources
PhD-level science reasoning
SurpassedDomain-expert validators score ~81%; frontier models are now near saturation at ~94%.
Sources
Competition mathematics
SurpassedStep-by-step competition math is largely solved at the frontier.
Sources
Real computer-use agents
At parityAgents reach approximate human parity on real desktop tasks — a 2026 crossing.
Sources
Novel abstraction & generalization
Live frontierCrosses the average human panel only at high compute and cost; the live general-reasoning frontier.
Research-grade mathematics
Not crossedResearch-grade math is far from solved; Tier 4 sits near 40%. The clearest remaining headroom.
Sources
The feedback loop
The exponential is driven by a compounding loop: AI tools accelerate the research and engineering that build the next model. These are the concrete, dated signals of that loop closing.
What this is and isn't. This is early recursive self-improvement — AI measurably accelerating AI R&D — not a model autonomously training its own successor end-to-end. Measured results and benchmarks are separated from forecasts below; forecasts are estimates, not evidence.
May 2026
StatementAnthropic put in writing that it sees AI contributing to speeding up the research and development of AI itself — an early recursive-self-improvement dynamic — and committed to publishing more on how AI tools accelerate its own work.
May 2026
DeploymentMost new code at Anthropic is authored by the model under human review — AI tooling materially inside the loop that builds the next model.
Nov 2025
BenchmarkOn seven ML research-engineering environments (calibrated against 71 eight-hour human-expert attempts), the best agents outscore human experts 4x when both are limited to 2 hours per task.
May 2025
Measured resultA Gemini-powered evolutionary coding agent discovered a 48-multiplication scheme for 4x4 complex matrix multiplication, improving on a result that had stood for 56 years.
May 2026
BenchmarkFrontier coding agents now resolve roughly four out of five real software-engineering issues from mature repositories, the substrate for AI maintaining its own toolchains.
Dec 2028
ForecastAnthropic's Jack Clark estimates a greater-than-even chance that, by end of 2028, a system could be told to make a better version of itself and do so autonomously. This is a forecast, not a measurement.
Methodology
The metric
The spine is the task-completion time horizon: the length of a task, measured in human-expert time, that a model completes at 50% reliability (METR's methodology). It is a single, behavioural number — not a benchmark score — which is why it tracks autonomy rather than recall.
The fit
We take log2 of each measured horizon and run a least-squares regression against the date. The slope is doublings per year; its inverse is the doubling time — ~5.2 months across the measured era, accelerating from the 7-month historical rate toward a recent ~4.5-month rate. The dashed line extends that fit; reference lines mark 1-day, 1-week, 1-month, and 1-year work horizons (8-hour work-days).
Uncertainty
Extrapolation is illustrative, not a date forecast. Stricter 80%-reliable horizons lag the 50% line, and a plausible compute-scaling slowdown pushes the 1-month horizon years later — which is why each milestone carries a wide band. The headline holds only as long as the recent doubling rate does.
Human baselines & RSI
Human-line baselines come from each benchmark's own expert panels. The recursive-self-improvement ledger separates measured results and benchmarks from forecasts; forecasts are labelled as such and never counted as evidence.
What this is not
This is not a probability that AGI or a singularity arrives by a given date, and not a claim of measured attainment. It is a structured, sourced way to read how fast autonomous capability is compounding. Source tiers and editorial policy follow the Landscape methodology.
Open data
License
The AGI Capability Frontier dataset is published by Scivity Labs under CC BY 4.0. You may reuse, remix, and republish with attribution.
Cite as
Scivity Labs (2026). AGI Capability Frontier — Time-Horizon Tracker [Dataset]. scivity.org/agi-progress
@misc{scivity_agi_capability_frontier_2026,
author = {{Scivity Labs}},
title = {AGI Capability Frontier --- Time-Horizon Tracker},
year = {2026},
howpublished = {\url{https://scivity.org/agi-progress}},
note = {Dataset. CC BY 4.0}
}