AI progress, checked and explained
See what AI can do, where people let it act, and whether the safeguards match the power it gets.
A new model set records on two AI tests in early September, and one of those tests is nearly used up. About one in five businesses in the United States report some use. Europe delayed its strictest AI rules. We still have much less evidence about whether safeguards work.
A source check finds updates. A human-readable summary changes only after its meaning and limits are reviewed. This edition re-checked 20 of 21 registered sources on Sep 7, 2026; the next review is due in four weeks. Read the method.
The picture in four parts
These charts tell four different parts of the story. They do not combine into one score.
What AI can do
Scores are rising on many tests, and one test built to resist memorization is nearly used up. But long jobs are chains, and one weak step can ruin the result.
What this cannot tell us: A rising test score does not tell us how close AI is to matching people at every task.
Epoch combines many tests into one score. Since fall 2024, its top score rose by about 14 points a year.
Keep in mind: The score is called the Epoch Capabilities Index, or ECI. It is not a percent, an IQ score, or a known distance to human-level AI.
Estimate Researchers calculated this from several observations.
The best reported setup perfectly finished about one in five of 108 long computer tasks.
Keep in mind: The same setup's average partial-credit score was 54.8%. That does not mean it finished the job.
Reported The source recorded a result, event, count, or public fact.
Where people let AI act
A system matters more when it can act in a workplace, move money, change code, or shape a decision.
What this cannot tell us: A use rate does not show how much access AI has inside each business.
Businesses with employees reported 22.4% use in late August and expected 25.9% within six months.
Keep in mind: The second number is what businesses expected, not a measured future result.
Reported The source recorded a result, event, count, or public fact.
Epoch estimates that five organizations hold 71% of the computing power used for AI.
Keep in mind: This estimate uses public evidence. Private infrastructure may be missing.
Estimate Researchers calculated this from several observations.
What has changed in the real world
Some jobs and industries are changing quickly. The whole economy still does not show one sudden break.
What this cannot tell us: More incident reports can mean more harm, better reporting, or both.
Stanford checked 12 measures. Seven looked normal, three showed mild change, and two showed strong change.
Keep in mind: Large changes in one job or industry can still be real. This view only asks whether the wider economy broke from its past pattern.
Estimate Researchers calculated this from several observations.
The AI Incident Database recorded 233 reports for 2024 and 362 for 2025.
Keep in mind: Reports cover different kinds of harm. The count is not the true incident rate.
Reported The source recorded a result, event, count, or public fact.
What can test, limit, or stop it
Companies are writing safety plans and Europe has passed a law, then delayed its strictest part. We have less public evidence that any of it works when a system is under pressure.
What this cannot tell us: Public disclosure is not the same thing as safety, but it lets outsiders check more.
Across 13 AI companies, researchers found about 41% of the public information they asked for.
Keep in mind: A low disclosure score does not by itself prove unsafe or unlawful behavior.
Reported The source recorded a result, event, count, or public fact.
Latest briefing
ARC-AGI-3 is a set of simple games with hidden rules. With one setup, Astra beat nearly every game about as efficiently as first-time human players. With the plain setup it scored 62.7%. The previous OpenAI model scored 7.8% in July.
GPT-6 Astra scored 169 on Epoch's index, first of 267 models. The top score in our August edition was 161.65.
In the two weeks to Aug. 23, 22.4% of businesses with employees said they used some AI, up from 21.5% in July. 25.9% expected to within six months.
15 signals, three lanes
Each lane starts with the plain answer. Open it only when you want the exact measure, source, and limit.
Evidence lane 01
Two records fell in September. Long jobs still break.
These sources cover test results, long technical tasks, and expert surveys. They do not add up to one score for human-like intelligence.
169 on the Epoch Capabilities Index, or ECI, with a 90% range of 165 to 174. The number only has meaning when models on the same scale are compared.
Estimate Researchers calculated this from several observations.
GPT-6 Astra had the highest combined score Epoch had measured when checked on Sep 7.
ECI is not a percentage, an IQ score, a reliability rate, or a known distance to AGI.
Epoch fits results from more than 50 benchmarks with an item-response model. Epoch's model page ranks GPT-6 Astra first of 267 models with a 90 percent interval of 165 to 174. The scale anchors Claude 3.5 Sonnet at 130 and GPT-5 at 150.
Sources: Epoch AI (opens in a new tab)
14 ECI points a year since September 2024, with a 90% range from 12 to 17 points.
Estimate Researchers calculated this from several observations.
Many included benchmark scores rose quickly during the fitted period.
The fit does not show uniform ability, durable progress, or a future rate.
Epoch fits a linear trend to state-of-the-art ECI scores since September 2024. Its 90 percent interval is plus 12 to plus 17 ECI per year; non-reasoning models rose 6 per year. The page header still reads Feb 5, 2026, so we cannot date the change from the 15.5 we recorded on Aug 18.
Sources: Epoch AI (opens in a new tab)
20.6% perfect completion and 54.8% average partial credit on OSWorld2.
Reported The source recorded a result, event, count, or public fact.
That setup completed about one fifth of the tested workflows perfectly.
The benchmark does not measure whole jobs, safe behavior, economic value, or every tool setup.
A Claude Opus 4.8 setup used batched tools for up to 500 steps on 108 long workflows.
11.98 expert-hours, with an estimated range from 5.28 to 60.56 hours. This describes task difficulty, not unattended runtime.
Estimate Researchers calculated this from several observations.
The model reached 50 percent predicted success near 12 hours of human task difficulty on this suite.
This is not 12 hours of unattended runtime or 12 hours of a normal job.
METR fits success against skilled human task time. The interval is 5.28 to 60.56 hours.
Sources: METR (opens in a new tab)
A fitted doubling time of 128.744 days on METR's stated task suite.
Estimate Researchers calculated this from several observations.
The horizon grew quickly on METR's software, machine learning, and cyber suite.
The trend need not continue, transfer to every field, or imply self-improvement. METR marks estimates above 16 hours as unreliable with this suite.
The post-2023 exponential fit has an interval of 104.428 to 158.012 days.
Sources: METR (opens in a new tab)
Survey medians of 2047 and 2116 among 2,778 researchers who had published at top AI venues.
Forecast This is a view of the future, not a current fact.
Many surveyed researchers expected very broad machine ability within decades.
These are not scientific deadlines. Answers varied widely and changed with the question.
A study surveyed 2,778 authors at leading AI venues, mainly in 2023, and reports medians across their probability forecasts.
Sources: Journal of Artificial Intelligence Research (opens in a new tab)
The report's assessment is conditional and does not assign a probability or deadline.
Assessment This is a source's judgment, not one direct measurement.
The reviewed evidence does not show that current systems are already superintelligent or beyond control.
The assessment does not make risk zero, prove a plateau, or supply a date. Experts disagree about progress through 2030.
The report reviews evaluations and expert views on capability, harmful behavior, and deployment access.
Sources: International AI Safety Report (opens in a new tab)
Evidence lane 02
Use edged up. The effects are still uneven.
These sources track business use, who controls the machines that run AI, changes in the economy, and reported incidents.
22.4% reported current use in Census period 107 (Aug 10 to 23). Respondents expected 25.9% within six months.
Reported The source recorded a result, event, count, or public fact.
AI use has spread to a meaningful share of United States employer businesses and rose slightly over the summer.
It does not show intensity, autonomy, productivity, or global use. A 2025 wording change starts a new series.
Census BTOS period 107 ran Aug 10 to 23. Standard errors are 0.36 and 0.47 percentage points. Period 106 (Jul 27 to Aug 9) read 21.8 and 25.9 percent. Period 108 had no published data on Sep 7.
An estimated 3.4 times annual growth rate in global AI computing capacity.
Estimate Researchers calculated this from several observations.
The estimated amount of available AI computing power rose quickly.
Not every chip is active, equally accessible, or turned one for one into ability.
Epoch combines revenue, company, and analyst evidence. The fitted interval is 3.2 to 3.7 times per year.
Sources: Epoch AI (opens in a new tab)
An estimated 71% share held by five large operators.
Estimate Researchers calculated this from several observations.
AI infrastructure is highly concentrated under this estimate.
Compute share is not a share of all AI decisions or outcomes. Private infrastructure may be missing.
Epoch estimates holdings from public financial and infrastructure evidence.
Sources: Epoch AI (opens in a new tab)
Seven measures showed no break from trend, three showed a mild break, and two showed a strong break.
Estimate Researchers calculated this from several observations.
The selected macro data did not yet show a clear economy-wide takeoff.
It does not rule out serious change for some workers or sectors, or a faster change later.
The July tracker fits trends for 12 economic indicators, models the residuals, and compares new values with bootstrapped predictive ranges. The research note shows an update dated Aug 10, 2026; the counts here are the July ones because the revised breakdown was not published in a form we could fetch.
362 reports for 2025, up from 233 for 2024.
Reported The source recorded a result, event, count, or public fact.
The count of documented reports rose.
Reports are not the true incident rate, have unequal severity, and do not show AI as the sole cause.
The AI Index summarizes public reports collected in a database that can revise records as evidence changes.
Sources: Stanford AI Index and AI Incident Database (opens in a new tab)
Evidence lane 03
Laws exist, and the strictest were pushed to 2027. We still know less about how well they work.
These sources track company safety plans, public disclosure, and laws. A written rule is not proof that a system stays safe under pressure.
Twelve companies with published or updated safety frameworks in 2025.
Reported The source recorded a result, event, count, or public fact.
Formal planning spread across frontier AI companies.
A published plan does not show that a company follows it or that its safeguards work. Scopes differ and most plans are voluntary.
The International AI Safety Report synthesizes public company framework activity.
Sources: International AI Safety Report (opens in a new tab)
An average transparency score of 40.69 out of 100.
Reported The source recorded a result, event, count, or public fact.
Public disclosure was limited under this audit.
A low score does not by itself prove unsafe or unlawful behavior. A changed rubric makes the 2024 comparison imperfect.
Researchers scored public company documents against a published 2025 rubric.
Sources: Stanford CRFM, Princeton University, and MIT (opens in a new tab)
General application and transparency duties began on Aug. 2, 2026. Annex III high-risk rules apply from Dec. 2, 2027 and Annex I rules from Aug. 2, 2028.
Reported The source recorded a result, event, count, or public fact.
Some AI duties are now enforceable in the European Union, and the high-risk duties were postponed before they took effect.
The law is not worldwide and does not by itself prove compliance or lower risk. The delay does not show whether risk rose or fell.
The European Commission timeline records the dates when parts of the law apply, as amended by Regulation (EU) 2026/1744.
Sources: European Commission AI Act Service Desk (opens in a new tab), European Commission (opens in a new tab)
Toward superintelligence?
A singularity would be a sudden jump in AI-driven change that becomes hard to predict or control. We do not know if it will happen. These are the signs we can check.
GPT-6 Astra set records on a combined test score and on ARC-AGI-3, a set of simple games with hidden rules. Long jobs and unfamiliar settings still expose simple failures.
Strong results across unfamiliar fields and messy real settings would make this sign more serious. Other researchers would need to repeat the result.
Some systems now handle harder technical tasks. That is not the same as finishing a normal workday without help, and the newest model has no published task-length number.
High success on long, messy work with little help and clear recovery from mistakes would change this status.
Some people and industries feel large effects, including workers in their early twenties in AI-exposed jobs. The wider economy still follows much of its old pattern.
A lasting jump in output, growth, or work that could not be explained another way would change this status.
Safety plans, outside checks, and laws are growing, but Europe delayed its high-risk rules by 16 months. Public proof that safeguards stop real failures is harder to find.
Repeated outside tests, clear incident records, limits that can be enforced, and proven recovery would change this status.
Choose your next step
Use screen stories, public claims, and daily AI life to learn the problem.
Keep upBriefingsRead the small set of changes that moved the evidence.
Learn the termsWikiBuild the concepts one plain-language page at a time.
Check the recordLibraryGo to papers, reports, provenance, and review status.