Today
17 h
Clears
2027
Task horizonBar 79 h
TOWARDSINGULARITY_
We track how far frontier AI has got in each domain, on that domain's own measure, with the gaps shown. A domain gets a date only when its data supports one, and no date here is an AGI forecast.
Software & ML research
Task horizon17 h
Doubling~4.1 mo
Arrival2027
Scientific reasoning
GPQA Diamond94.8%
Doubling~6.7 mo
Arrival2027
Advanced mathematics
FrontierMath T487.8%
Doubling~2.0 mo
Arrival2026
Agentic computer use
No measurementPublished elsewhere, not wired here yet.
Robotic manipulation
No measurementPublished elsewhere, not wired here yet.
Self-driving
No measurementPublished elsewhere, not wired here yet.
Video understanding
Metric N/ANot a gap to be filled. Needs a different instrument, not a missing dataset.
Medicine
No measurementNo dated, independently maintained measurement exists. This site will not invent one.
Law
No measurementNo dated, independently maintained measurement exists. This site will not invent one.
Finance
No measurementNo dated, independently maintained measurement exists. This site will not invent one.
Domain coverage
3 of 10 domains carry a date
Today
94.8%
Clears
2027
GPQA DiamondBar 99.0%
Today
87.8%
Clears
2026
FrontierMath T4Bar 90.0%
Within one test run's noise of the bar.
No dated series yet for the other 7: Agentic computer use, Robotic manipulation, Self-driving, Video understanding, Medicine, Law, Finance.
Why a full board
still is not a date.
Filling every row would raise the floor, not settle the question. Two reasons a complete board would still stop short of a date.
The slowest domain decides
If generality is the claim, the right aggregation is a minimum, not a mean.
A system superhuman at code and helpless at everything else is not general, and an average would hide exactly that. So an AGI date would be set by the slowest domain on the board, and the slowest domains are the ones with the least data. Adding more well-measured coding benchmarks would not move it at all.
A mean rides the one tall bar. The floor is set by the domains nobody has measured.
Breadth is not a domain
Coverage is necessary here, not sufficient.
Most definitions of AGI require transferring to problems nobody enumerated, and a checklist of domains cannot test for that: you can clear any fixed list by building one specialist per domain. Even a full board would not settle it.
Every box ticked, and the list still does not close on the right.
How fast is frontier AI
moving?
Four separate measurements, each in its own unit, each tracked publicly by someone else, last read on October 6, 2026. They are not combined into a score: there is no defensible scale on which a training run and a price are the same kind of quantity, and inventing one would smuggle in a weighting nobody could argue with.
Compute
1026.70 FLOP
Largest disclosed training run
Source: Epoch AI
Cost
$0.090/Mtok
gemma-4-31b-it
Cheapest inference clearing a capability floor
Source: OpenRouter
Benchmarks
1508 Elo
claude-fable-5
Top rating, style-controlled leaderboard
Source: LMArena
Forecasts
38% by 2030
market-implied, horizon 2030
Market-implied probability by the horizon year
Source: Manifold
From cited data
to a dated trend.
Read the
cited sources
Training compute from Epoch AI, task horizons from METR, test scores from Epoch AI’s benchmarking hub, and cost, rating and market readings from OpenRouter, LMArena and Manifold. Every input traces back to its source.
Report each
in its own unit
No shared scale and no composite. The four readings are published as they come, a FLOP count, a price, a rating and a probability, because there is no defensible axis on which those are the same kind of quantity.
Publish & score
the predictions
A measured trend per domain, on that domain’s own measure, with a dated forecast where the data supports one, and public predictions graded true or false as they come due.
The tools moving
the numbers.
One standout per layer of the stack: assistant, coding, research, agents, media and safety. Picked by the editors, and none of it feeds a calculation on this site.
OpenAI · Assistant
ChatGPT
The mass-market on-ramp to frontier models: a general assistant for writing, reasoning and code.
Visit siteAnysphere · Coding
Cursor
AI-native code editor that edits across an entire repository.
Visit sitePerplexity AI · Research
Perplexity
Answer engine that cites its sources as it searches the live web.
Visit siteCognition · Agents
Devin
Autonomous software agent that plans and ships multi-step tasks.
Visit site01 / 01
Every input traces back to a cited source
When a number needs
a thousand words.
Methodology
What these forecasts do not measure
The dated forecasts are not about AGI. Each one is about a single measurement in a single domain. What those numbers cover, and why raising the bar does not widen them.
Oct 2026 · 5 min read
Methodology
Why the forecast gives no single date
A point estimate hides everything that matters. The case for publishing a distribution, and how this one is built.
Oct 2026 · 4 min read
Data
Training compute is still doubling about every five months
The clearest leading indicator hasn't bent, and why it still does not move a single date on this site.
Oct 2026 · 4 min read
01 / 01