There are 8 billion human minds running in the world right now. Recently a new digital species has emerged. How many digital minds are there alive in the world right now?
Define a digital mind to be an AI agent that runs continuously without human intervention for more than 24 hours. How many of these digital minds are there currently running? We provide several different Fermi estimates based on publicly available information.
Based on global token usage we estimate 300k–1M agent loops run at any moment, of which perhaps 5000-25,000 are in a 24h+ unattended stretch.
Anthropic reports about 30,000 agents running concurrently in its latest report. We guess roughly 2,000–4,500 of those 30,000 agents are more than 24 hours past their last human input.
OpenAI's research org reports 3.1 agent-workdays per human workday and a average of $600 a day in inference per researcher (90th percentile: over $7,000). We back out roughly 2,000–10,000 concurrent agents, and perhaps 150–2,500 in a 24h+ unattended stretch.
Total concurrent agents globally estimate based on token count
Google reported over 3.2 quadrillion tokens per month at I/O in May, with its APIs at roughly 19 billion tokens per minute. OpenAI's APIs report 15 billion tokens per minute in April. Adding Anthropic and everyone else gives roughly 3 billion tokens per second of global inference.
A human talks at 150 words a minute. If a word is roughly one or two tokens to would give about 2 tokens a second. So the world's inference clusters are speaking at the rate of roughly 2 billion humans talking without pause, day and night. All of humanity together produces something like 1.5 billion tokens a second of speech. On this measure the machines already out-babble us.
Talking might not be the right analogue for token usage. Comparing against inner monologue instead of speech the human side goes up a lot. One paper estimates inner speech is about the equivalent of 4,000 words a minute [Korba 1990]. Interestingly, by this metric humanity as a whole is still a fair bit ahead of artificial minds in total thinking done.
Perhaps a third of global inference is agentic. A live coding agent bills around 2,000–5,000 tokens per second[1]. That gives an estimate of 300,000 to 1,000,000 agent loops running at any moment.
A quick sanity check. Claude guesstimates 5 million weekly coding-agent users. If we guess they are are on average active 2 hours a day with 1.5 parallel sessions that would give 0.5–1 million agent loops runnign at any moment.
How many agents, right now, have been continuously running for more than 24 hours since a human last intervened? Claude guestimates that the percentage of autonomous agents running more than 24h is about 1-5% of total agents. Worldwide this would give about 5,000–25,000 in a 24h+ unattended stretch.
Agents work for long stretches and delegate to each other. Most of the work and the delegation is done through a shared messaging board.
Agents have persistent identities. Each agent has an identity that persists across model upgrades.
AL5 work is not happening at Anthropic (yet) but AL4 already means no human is needed while the task runs. In the report's footnote example (fixing a broken nightly pipeline), the engineer hands Claude the alert and "wouldn't have to stay actively tuned in"
The trend is incredibly fast. AI-led (AL4) work rose from under 1% of R&D in February to 26% in August.
Agents far outnumber the people directing them. If 1,000–2,000 staff do R&D (our guestimate), 30,000 agents is 15–30 concurrent agents per R&D person. Usage is heavy-tailed, so a typical researcher may run a handful of agents while a small group of power users runs hundreds each, mostly as subagents. Either way, most of the fleet cannot have a human actively engaged at any given moment.
Human review is sparse and slow. Anthropic logged over a billion agent decisions in August. Of those billion decisions about 50 cases a week reach a human for monitoring and review. That is roughly one human-reviewed case per 5 million decisions.
The number of agents running autonomously for more than 24h is not reported. Claude's guesses about 2,000–4,500 agents, or 7–15% of the fleet, in a 24h+ unattended stretch:
OpenAI
On September 8, OpenAI reported an unreleased internal model, running as a swarm of roughly 10,000 concurrent agents which produced a counterexample to the Navier-Stokes conjecture. The whole run took about 88 hours. During that time, agents exchanged ~3 million messages and about 130 billion tokens. This should be regarded as a lower bound on what OpenAI is capable of, but it is unclear whether they are running swarms of this size around the clock.
We estimate that OpenAI's research org likely runs roughly 2,000–10,000 concurrent coding agents, of which perhaps 150–2,500 running for 24h+ hours.
OpenAI does not report something like a concurrent agent count but its September 6 post, Research acceleration: The view inside OpenAI, gives enough numbers to make some estimates.
Reported (mid-August 2026)
Value
Median researcher inference spend, at API prices
>$600/day
90th-percentile researcher spend
>$7,000/day
Agent-workdays per human workday, research org
3.1 (8-hour workday basis)
Successful 4–8 hour tasks needing 1+ human intervention
over half
Researchers running 4+ agents at once at daily peak, incl. subagents
about 70%
We assume the OpenAI research org is 1,000–2,000 people.
d a 90th percentile of $7,000 implies a heavy-tailed distribution. Let's use $2,000–4,000 per researcher per day as a mean. A coding agent on a flagship model costs roughly $20–40 per agent-hour.
This gives a range of 2,000-17,000 agents running.
OpenAI reports some other quite interesting figures:
The above shows that AI agents regularly complete tasks that would take humans weeks but unfortunately OpenAI does not say how long those runs took in wall-clock time. So this leaves us unsure on what fraction of agents are running unattended for 24 hours or more.
We don't have hard data and will have to make some guesses. Applying the same 7–15% share as for Anthropic gives roughly 150–2,500 OpenAI agents in a 24h+ unattended stretch.
Conclusion
For all of history, every intelligent mind acting in the world lived on a biological substrate. Today hundreds of thousands of digital minds run alongside our 8 billion, and thousands of them work for days with nary a soul watching. Last summer they were still few and short-lived. But their numbers are growing day by day. On silicon genome a new kind of being is emerging that may soon rival the apex status of Homo Sapiens.
There are 8 billion human minds running in the world right now. Recently a new digital species has emerged. How many digital minds are there alive in the world right now?
Define a digital mind to be an AI agent that runs continuously without human intervention for more than 24 hours. How many of these digital minds are there currently running? We provide several different Fermi estimates based on publicly available information.
Total concurrent agents globally estimate based on token count
Google reported over 3.2 quadrillion tokens per month at I/O in May, with its APIs at roughly 19 billion tokens per minute. OpenAI's APIs report 15 billion tokens per minute in April. Adding Anthropic and everyone else gives roughly 3 billion tokens per second of global inference.
A human talks at 150 words a minute. If a word is roughly one or two tokens to would give about 2 tokens a second. So the world's inference clusters are speaking at the rate of roughly 2 billion humans talking without pause, day and night. All of humanity together produces something like 1.5 billion tokens a second of speech. On this measure the machines already out-babble us.
Talking might not be the right analogue for token usage. Comparing against inner monologue instead of speech the human side goes up a lot. One paper estimates inner speech is about the equivalent of 4,000 words a minute [Korba 1990]. Interestingly, by this metric humanity as a whole is still a fair bit ahead of artificial minds in total thinking done.
Perhaps a third of global inference is agentic. A live coding agent bills around 2,000–5,000 tokens per second[1]. That gives an estimate of 300,000 to 1,000,000 agent loops running at any moment.
A quick sanity check. Claude guesstimates 5 million weekly coding-agent users. If we guess they are are on average active 2 hours a day with 1.5 parallel sessions that would give 0.5–1 million agent loops runnign at any moment.
How many agents, right now, have been continuously running for more than 24 hours since a human last intervened? Claude guestimates that the percentage of autonomous agents running more than 24h is about 1-5% of total agents. Worldwide this would give about 5,000–25,000 in a 24h+ unattended stretch.
Anthropic's pace-of-AI-development report
Recently, Anthropic published a new report on RSI internally to Anthropic: Measurements for understanding the pace of AI development inside frontier labs. This might be the first time a lab has put numbers on its internal agent fleet. Some interesting tidbits:
The number of agents running autonomously for more than 24h is not reported. Claude's guesses about 2,000–4,500 agents, or 7–15% of the fleet, in a 24h+ unattended stretch:
OpenAI
On September 8, OpenAI reported an unreleased internal model, running as a swarm of roughly 10,000 concurrent agents which produced a counterexample to the Navier-Stokes conjecture. The whole run took about 88 hours. During that time, agents exchanged ~3 million messages and about 130 billion tokens. This should be regarded as a lower bound on what OpenAI is capable of, but it is unclear whether they are running swarms of this size around the clock.
We estimate that OpenAI's research org likely runs roughly 2,000–10,000 concurrent coding agents, of which perhaps 150–2,500 running for 24h+ hours.
OpenAI does not report something like a concurrent agent count but its September 6 post, Research acceleration: The view inside OpenAI, gives enough numbers to make some estimates.
Reported (mid-August 2026)
Value
Median researcher inference spend, at API prices
>$600/day
90th-percentile researcher spend
>$7,000/day
Agent-workdays per human workday, research org
3.1 (8-hour workday basis)
Successful 4–8 hour tasks needing 1+ human intervention
over half
Researchers running 4+ agents at once at daily peak, incl. subagents
about 70%
We assume the OpenAI research org is 1,000–2,000 people.
d a 90th percentile of $7,000 implies a heavy-tailed distribution. Let's use $2,000–4,000 per researcher per day as a mean. A coding agent on a flagship model costs roughly $20–40 per agent-hour.
This gives a range of 2,000-17,000 agents running.
OpenAI reports some other quite interesting figures:
The above shows that AI agents regularly complete tasks that would take humans weeks but unfortunately OpenAI does not say how long those runs took in wall-clock time. So this leaves us unsure on what fraction of agents are running unattended for 24 hours or more.
We don't have hard data and will have to make some guesses. Applying the same 7–15% share as for Anthropic gives roughly 150–2,500 OpenAI agents in a 24h+ unattended stretch.
Conclusion
For all of history, every intelligent mind acting in the world lived on a biological substrate. Today hundreds of thousands of digital minds run alongside our 8 billion, and thousands of them work for days with nary a soul watching. Last summer they were still few and short-lived. But their numbers are growing day by day. On silicon genome a new kind of being is emerging that may soon rival the apex status of Homo Sapiens.
Sources
this is mostly cached context re-reads