Last Updated: November 5, 2026
The 36-Month Miracle
⚠ June 2026 Update — The Miracle Went Autonomous
When this page was written, the equation read "intelligence became cheap." Now it also works on its own and improves itself. In April 2026, Anthropic launched Managed Agents. Around then, a Rakuten engineer had Claude Code build a feature across a huge codebase on its own:
12.5M lines in the codebase
7 hours in a single autonomous run
99.9% accuracy
METR's autonomy benchmark found Claude Opus 4.6 (Feb 2026) could finish, half the time, tasks that take a person about 12 hours. That length has been doubling every 4–7 months. In May, METR put an early Claude Mythos Preview at likely at least 16 hours, the most its tests can reliably measure. METR stopped updating the page in September 2026.
Jack Clark (Anthropic policy lead) put the odds at 60%+ that a top AI trains its own successor with no human involved by the end of 2028, in his Import AI newsletter of May 4, 2026.
Anthropic's own timeline (per Redwood Research's analysis of their OSTP filing) projects AI making research engineers more than 10x faster by the end of 2026, and possibly fully automating AI research by early 2027.
The multipliers below were already huge. Now they're the low estimate. Companion piece: AI Is Getting Really Serious.
July 2026 Edit: The coding test this page leans on is wearing out. OpenAI itself now says
SWE-bench Verified "no longer measures frontier coding capability": top models score 88–95%, leaving little room to tell them apart. In July 2026, Fable 5 led at 95.0% Verified / 80.3% Pro, ahead of the Opus 4.8 numbers cited below. Anthropic has since released Fable 5.1. A test maxing out while AI keeps improving backs up this page's point.
Intelligence used to be the most expensive AND rarest commodity on the planet.
In 3 years, we turned it into a utility cheaper than tap water.
In early 2023, AI was about as capable as a smart high-schooler. By 2025, it was producing Post-Doc research.
Then came agents, digital employees that act on their own.
Capability
60%
smart high-schooler
×
Speed
3x
heavy human review
×
=
Less capable AI needs more human checking and more redo passes before its work is usable, which eats into the speed gain.
Near expert level, checking barely matters. Real costs vary; this is simplified to show the trend.
First
The "Inbox Man"
Data entry, routing, routine compliance. It's just inputs, rules and outputs, with nobody needed in person, so it goes instantly. AI does it better and never asks for vacation.
Slower
Physical Work
Code moves at light speed, but moving boxes takes a body. Trades and physical work last longer because they need someone on site.
BUT...
Last
The "Human Shield"
CEOs get paid for status, not logic. They're the person the tribe trusts, the one the law can sue, the face of "us." They become Wise Idea Coordinators.
July 2026 Edit: The "digital employees" line above isn't a metaphor anymore. Salesforce's Agentforce numbers show real companies paying real money for agents doing "Inbox Man" work today.
$800M Agentforce ARR
+169% year over year
18,500+ customers
Moonshot AI's Kimi 2.5 also shipped an Agent Swarm mode, where the model was trained to decide for itself when to start helper agents, with no human running them. The "Wise Idea Coordinator" role above was meant for the last jobs to go. It's now the job for anyone still working a desk job.
The Economic Feedback Loop
Spreadsheet Logic
"Lay off 50% of staff to maximize quarterly margins."
Social Reality
50% unemployment = no customers. The company dies with the community.
The Abundance Question
The Vision
AI creates so much wealth that everyone benefits through redistribution.
Structural Reality
Cash doesn't help if housing, energy, healthcare remain scarce. Abundance = access.
"It's as if someone were out there making up pointless jobs just for the sake of keeping us all working."
David Graeber, "On the Phenomenon of Bullshit Jobs," Strike! magazine, August 2013
AI removes the bullshit. Instead of leaving us idle, it re-tasks us.
Cleaning the oceans and land
Retrofitting homes for energy independence
Rebuilding infrastructure
High-touch elder care
Community mental health
Regenerative agriculture
No government program involved: this is how people survive. Hover to see how.
About the author
I'm Scott Covert. I'm 60, based in Peterborough, Ontario, and I've been using the top AI models for real work — research, writing, business systems, medical advocacy, financial analysis — since early 2023. I built AI Over 50 for adults my age who'd benefit from these tools more than anyone but get the worst instruction.
The numbers on this page are mostly public benchmark results (METR, SWE-bench, ARC-AGI) and direct quotes from frontier-lab announcements. A few are other people's forecasts, such as Jack Clark's odds of 60%+ by 2028. Every claim is sourced below. If you find one that doesn't hold up, email me: [email protected] — I update when I'm wrong.
More work: Substack · Medium · AI Over 50
Sources cited on this page
-
Anthropic Managed Agents (April 2026) — the-decoder.com — cloud-hosted autonomous agents; plus Claude Code's Rakuten 12.5M-line, single 7-hour run.
-
METR Autonomy Time Horizons — metr.org/time-horizons — Claude Opus 4.6 measured at ~12-hour 50% task horizon (Feb 2026); horizon doubling every 4–7 months.
-
Jack Clark, Anthropic — Import AI 455 (May 4, 2026) — jack-clark.net — "a likely chance (60%+) that no-human-involved AI R&D – an AI system powerful enough that it could plausibly autonomously build its own successor – happens by the end of 2028." Axios coverage, May 7.
-
Anthropic OSTP filing — Redwood Research analysis — blog.redwoodresearch.org — mid-2026 ~3x research engineer acceleration; end of 2026 >10x; possible full AI R&D automation early 2027.
-
SWE-bench Verified leaderboards (June 2026) — swebench.com — Claude Mythos Preview 93.9%, Claude Opus 4.8 88.6%, Claude Opus 4.7 87.6%.
-
Claude Opus 4.7 announcement (April 16, 2026) — anthropic.com/news/claude-opus-4-7 — SWE-bench Pro 64.3%, GPQA Diamond 94.2%.
-
OpenAI GPT-5.5 ("Spud") launch (April 23, 2026) — openai.com — Terminal-Bench 2.0 82.7%, ARC-AGI-2 85.0%, FrontierMath T1-3 51.7%.
-
SWE-bench leaderboard, July 2026 — swebench.com — Fable 5 led at 95.0% Verified / 80.3% Pro; OpenAI has publicly stated the benchmark "no longer measures frontier coding capability."
-
Salesforce Agentforce & Moonshot Kimi 2.5 (July 2026) — Salesforce reported $800M in Agentforce ARR (+169% YoY, 18,500+ customers); Kimi 2.5's Agent Swarm mode is the first model RL-trained to decide when to launch sub-agents itself.
Note: the rotating equation above (capability × speed × scale) is a rough way to show how the gains multiply, not a measured benchmark. The sources above back the rest of the page.
The spreadsheet is being rewritten,
and you decide what it all gets used for.