AI Clock

AI is scaling. Watch it happen.

Real-time metrics on AI growth — tokens, spending, energy, compute, and research — from public data sources.

Since you arrived — ago
—
messages sent
—
spent on chips
—
revenue earned
—
autonomy gained
When's the next frontier model?
AnthropicClaude Opus 5.5
73%~7d
~Oct 2026
OpenAIGPT-6 Astra
63%~23d
~Nov 2026
GoogleGemini 3.8 Flash
107%3d overdue
~Oct 2026

How are AI models improving on real tasks?

SOTA scores on key benchmarks with trend-based saturation estimates. Source: Epoch AI (CC-BY).Expand all

Software Engineering
SWE-bench Verified
83.5%
Claude Opus 4.7
Mathematics
FrontierMath
93.7%
GPT-6 Astra
IT & DevOps
TerminalBench
84.7%
GPT-5.5
Reasoning
SimpleBench
88.4%
Claude Opus 5.5
Expert Knowledge
Humanity's Last Exam
54.8%
GPT-6 Astra
Professional Work
APEX-Agents
82.2%
gemini-4-argon
Abstract Reasoning
ARC-AGI-2
93.3%
GPT-6 Astra
Computer Use
OSWorld
86.0%
Claude Fable 5
Abstract Reasoning
ARC-AGI-3
62.7%
GPT-6 Astra
Multimodal Reasoning
MMMU-Pro
83.6%
Gemini 3.5 Flash
Web Research
BrowseComp
91.5%
GPT-6 Astra

How autonomous is AI?

AI reliably solves tasks that take a human
hours(Claude Mythos)
Succeeds 80% of attempts on software, cybersecurity & ML tasks.
doubling every ~107 days
Data: Aug 2026
AI solves half the tasks that take a human
hours(Claude Mythos)
Succeeds 50% of attempts. In 2023, GPT-4 managed just 4 min.
doubling every ~107 days
Data: Aug 2026
80% success rate
hours
95% CI – hrs
50% success rate
hours
95% CI – hrs
until AI gains >1s of autonomy per real second
80% reliable — ~Apr 2029
—
rate: 0.003s/s · doubling ~107d
95% CI: Nov 2028 – Sept 2029
50% median — ~Jul 2028
—
rate: 0.016s/s · doubling ~107d
95% CI: Mar 2028 – Nov 2028
In 2023, AI could solve half the tasks that take a human 4 minutes.
Now it handles tasks up to 17.4 hrs.*
Last published: Claude Mythos (Apr 2026). Autonomous task horizon doubling every ~107 days (95% CI 94–121) (2024+ trend). Source: METR Horizon v1.1 benchmark.
* Software engineering, cybersecurity, and machine learning tasks only.
Show chart
2023–2028
SOTA frontierNon-SOTA2024+ fit (2x / ~107d)Forward projection (from last published)All-time (2x / ~176d)95% CI
Top 5 by horizon
#Modelp50 HorizonDate
1Claude Mythos17.4 hrsApr 2026
2Opus 4.612.0 hrsFeb 2026
3GPT-5.25.9 hrsDec 2025
4Opus 4.54.9 hrsNov 2025
5Gemini 3 Pro3.7 hrsNov 2025
p50 horizon = the longest task, by human completion time, that AI solves 50% of the time. From 4 min (GPT-4) to 17.4 hrs (Claude Mythos) in under 3 years.
See every model on the timeline →

How cheap is autonomous AI getting?

AI Clock metric · cheapest frontier-tier
$5.75
89% success on Verified
to make a frontier AI fix one bug
currently GPT-5.5 · SWE-bench Verified
A human engineer needs up to an hour. The AI bills you less than a sandwich.
As low as $0.76 with open-weight (DeepSeek V3.2) — but 61% success, so it's a different trade.
● Mar 2023: $433 (GPT-4)● today: $5.75 (GPT-5.5)
Frontier model trajectory
#ModelCostSuccessDate
1GPT-5.5$5.7589%Apr 2026
2Claude Opus 4.7$10.388%Apr 2026
3Claude Opus 4$31.073%May 2025
4Gemini 2.5 Pro$4.1164%Apr 2025
5Claude 3.5 Sonnet (new)$4.4449%Oct 2024
6GPT-4o$7.4833%May 2024
7Claude 3 Opus + SWE-agent$60.511%Apr 2024
8GPT-4 + SWE-agent$20.818%Apr 2024
9GPT-4 (8K)$4332%Mar 2023
AI has gone from $433 per resolved task to $5.75 — 75× cheaper in 3 years. Cost halves every ~251 days (95% CI 127–2244d).

How is AI changing the economy?

this year
invested in AI companies
More than global VC investment in all of biotech.
+$28,473 / second
▲ collapse chart
Data: Jun 2026
Sep 2026
23.8%
of US businesses now use AI
Faster than PC adoption was at the same stage.
↑ from 3.7% in 2023
Census Bureau BTOS — updates biweekly
Data: Sep 2026
▲ collapse chart
Aug 2026
6.7%
of US job postings mention AI
54x growth in two years. From 1 in 1,000 to 1 in 19 postings.
↑ from 2.1% in Q1 2024
Indeed Hiring Lab — updates monthly
Data: Sep 2026
▲ collapse chart
this year
US layoffs attributed to AI
101,743 AI-attributed cuts in the first six months of 2026 alone.
~477 per day
*Ticks roughly every 10 minutes — Challenger monthly reports
▲ collapse chart
Data: Sep 2026

How much is AI being used right now?

today
AI messages sent
Over two messages per person on Earth. Every single day.
+322,951 every second
Data: Sep 2026
today
AI tokens generated
Enough text to fill the Library of Congress dozens of times over. Every day.
+1,087,962,963 per second
▲ collapse chart
Data: Sep 2026
Users
Meta AIMAU
ChatGPTMAU
GeminiMAU
ClaudeMAU
GrokMAU
DeepSeekMAU
PerplexityMAU
All figures monthly active users (MAU). ChatGPT estimated from WAU.

What's powering all of this?

this year
spent on AI chips
Doubling roughly every 9 months. Was $32B in 2023.
+$14,577 / second
▲ collapse chart
Data: Apr 2026
today
MWh
consumed by AI data centres
14 GW of continuous draw — over a dozen nuclear reactors running non-stop.
+4.0 MWh / second
Data: Apr 2026
today
tonnes
CO2 from AI data centres
1.9 tonnes every second. Based on a grid average of 0.49 kg CO₂/kWh.
+2.0 tonnes / second
Data: Apr 2026
cumulative
AI chips deployed worldwide
Each chip costs $25,000–$40,000. And there's a global waitlist.
+9.9 million this year
Data: Apr 2026

How good and cheap are the models?

median price
$0.40
per million tokens
Falling 6x faster than solar panels did.
↓ 84% since 2024
Snapshot — drops episodically when new models launch
Data: Jun 2026
▲ collapse chart
frontier compute
1,000 trillion trillion
FLOP for frontier model training
More compute in one training run than all of CERN's 2020 data processing.
↑ 20x since 2023
Snapshot — jumps when a new frontier model trains
Data: Sep 2026
▲ collapse chart

How fast is AI research moving?

this year
AI papers on arXiv
More AI papers than physics, math, or biology papers on arXiv.
+472 per day
Data: Sep 2026
today
new models on HuggingFace
Under 25 seconds between uploads. The long tail of AI development.
~3,639 per day
Data: Jun 2026
NeurIPS 2025
5,823
accepted papers
NeurIPS alone now publishes more papers than most journals.
↑ 476% since 2018
Annual conference count — next update Dec 2026
Data: Aug 2026
this year
AI patents filed (USPTO)
11x growth in AI patent filings over the last decade.
~43 per day
Data: Feb 2026

Where's the money going?

this year
AI company revenue earned
From essentially $0 three years ago.
+$4,962 / second
▲ collapse chart
Data: Sep 2026
Revenue by company (this year)
Anthropic+$3,169/s
OpenAI+$1,584/s
Z.ai (Zhipu)+$51/s
DeepSeek+$32/s
Mistral AI+$32/s
Moonshot AI+$32/s
MiniMax+$25/s
xAI+$16/s
Perplexity+$14/s
Cohere+$8/s

What's the big picture?

2025
38%
of all arXiv papers are AI
From 1 in 6 papers to nearly 2 in 5 — in seven years.
↑ from 16% in 2018
Annual share — see AI papers counter above for real-time
Data: Sep 2026
▲ collapse chart
Time travel
Share
Embed
Send feedback