AI Clock

AI is scaling. Watch it happen.

Since you arrived — ago—— messages · — on chips · — revenue
today so far
 
messages sent
 
tokens generated
 
revenue earned
An AI agent can work autonomously for
hoursClaude Opus 4.6
In 2023, GPT-4 managed 3.5 minutes. Now: 14.5 hours.
doubling every ~100 days
projected right now
seconds
≈ hours
In 2023, AI could work alone for 3 minutes.
Now it handles 14-hour tasks.
Autonomous task horizon doubling every ~100 days (2024+ trend). Source: METR Horizon v1.1 benchmark.
Mar 2026PROJECTED
17.7 hrs
?
projected p50 horizon (2024+ trend, 2x / ~100 days)
2.0 hrs4.0 hrs8.0 hrs14.5 hrs1.0 days2.0 days4.0 days7.0 daysAdd auth to an existing app2 hoursBuild and deploy a REST API4 hoursReplace a junior dev for a day8 hoursPrototype an MVP from scratch14.5 hoursRun a full QA cycle solo1 dayBuild a complete SaaS app2 daysMigrate a legacy codebase4 daysReplace a small dev team for a week1 week?17.7 hrs

AI could handle (projected)

Autocomplete a sentence2.4 seconds
Answer a question30 seconds
Write a short email1 minute
Fix a bug from a stack trace3.5 minutes
Write tests for a module10 minutes
Review and refactor a PR30 minutes
Build a landing page from a design1 hour
Add auth to an existing app2 hours
Build and deploy a REST API4 hours
Replace a junior dev for a day8 hours
Prototype an MVP from scratch14.5 hours

Not yet

Run a full QA cycle solo1 day
Build a complete SaaS app2 days
2023–2028
SOTA frontierNon-SOTA2024+ trend (2x / ~100d)All-time (2x / ~176d)95% CI
ModelReleasedp50 Horizon% Tasks Solved
Claude Opus 4.6Feb 202614.5 hours79%
GPT-5.2Dec 20256.6 hours75%
GPT-5.3 CodexFeb 20266.5 hours75%
Claude Opus 4.5Nov 20255.3 hours73%
Gemini 3 ProNov 20253.9 hours71%
GPT-5Aug 20253.6 hours69%
o3Apr 20252 hours64%
o1Dec 202438 min51%
Claude 3.5 SonnetOct 202420 min45%
GPT-4(Mar 2023)Mar 20233.5 min29%
p50 horizon = task duration where model succeeds 50% of the time. % Tasks Solved = share of METR benchmark tasks completed at the p50 threshold. From 3.5 minutes (GPT-4) to 14.5 hours (Opus 4.6) in under 3 years — a 250x improvement.

Usage

today
AI messages sent
40,000 messages every second — 24/7/365.
+40,104 every second
today
AI tokens generated
Google alone generates 40 trillion tokens per day.
+800,011,574 per second
click for chart
Users
Google AI OverviewsMAU
ChatGPTMAU
Meta AIMAU
GeminiMAU
ClaudeMAU
DeepSeekMAU
GrokMAU
PerplexityMAU
All figures monthly active users (MAU). ChatGPT estimated from WAU.

Infrastructure

this year
spent on AI chips
That's $369 million per day on AI hardware.
+$4,269 / second
click for chart
today
MWh consumed by AI compute
91.1 TWh/year. More than Belgium.
+2.9 MWh / second
today
tonnes CO2 from AI compute
Like 9.7 million cars on the road. More than Finland.
+1.4 tonnes / second
cumulative
AI chips deployed worldwide
786 GPU clusters. 13.3 GW total power capacity.
+7.1 million this year

Capability

median price
$0.40
per million tokens
API costs halve roughly every 6 months.
↓ 84% since 2024
Snapshot — drops episodically when new models launch
click for chart
frontier compute
5×1026
FLOP for frontier model training
Doubling roughly every 6 months.
↑ 13x since 2023
Snapshot — jumps when a new frontier model trains
click for chart

Research

this year
AI papers on arXiv
More AI papers than physics, math, or biology papers on arXiv.
+810 per day
today
new models on HuggingFace
3,600+ models uploaded daily. Most are fine-tunes.
~3,639 per day
NeurIPS 2024
4,561
accepted papers
Was 1,011 in 2018. Growing 28.5% per year.
↑ 351% since 2018
Annual conference count — next update Dec 2025
this year
AI patents filed (USPTO)
More AI patents filed per year than total biotech patents in 2010.
~43 per day

Money

this year
AI company revenue earned
More than Hollywood's global box office. In API fees alone.
+$1,112 / second
click for chart
Revenue by company (this year)
Google Cloud*+$1,775/s
OpenAI+$634/s
Anthropic+$444/s
xAI+$16/s
Mistral AI+$13/s
Perplexity+$5/s
*Google Cloud is total cloud revenue (includes Gemini but also non-AI services). Standalone Gemini revenue not disclosed.

Signals

2025
42%
of all arXiv papers are AI
From 1 in 5 papers to nearly 1 in 2 — in seven years.
↑ from 18% in 2018
Annual share — see AI papers counter above for real-time
click for chart
Time travel
Share
Send feedback