SWE-bench Verified (one year)
60% → ~100%
Software engineering benchmark capability, nearly saturated.
RESEARCH BRIEF / RB-02 · V2.1
RB-02 · AI After
The Gold Rush
When Intelligence Becomes Cheap,
What Becomes Scarce?
From AI Access to AI Leverage
What Becomes Scarce?
From AI Access to AI Leverage
AI is making intelligence abundant.
The ability to turn intelligence into reliable outcomes is becoming scarce.
The Question
What becomes scarce when intelligence becomes cheap?
Executive Summary
Three things, true at the same time.
As of August 2026, the evidence is clear: AI has gone mainstream, capability continues to accelerate — but the harder problem is turning that capability into measurable, reliable economic value.
01
AI has gone mainstream
88% of organizations use AI; 70% use generative AI in at least one business function. Population adoption reached 53% in three years — faster than the PC or the internet.
Stanford AI Index 2026
02
AI capability is accelerating
SWE-bench Verified went from ~60% to near 100% in a single year. GPT-5.6, Claude Sonnet 5 / Opus 5 have pushed capability toward execution.
Stanford AI Index 2026 · OpenAI · Anthropic
03
But value capture is still hard
In 2025, METR measured experienced developers taking 19% longer on tasks. By 2026, because developers refuse to work without AI, reliable measurement is becoming increasingly difficult. New data points toward gains, but estimates remain uncertain and vary across studies.
METR 2025 RCT · METR Feb 2026 · Management Science 2026
Main Thesis
When intelligence becomes cheap, reliable execution becomes scarce.
When intelligence becomes cheap, reliable execution becomes scarce.
An analytical position supported by the evidence below, not a single measured finding
01 — The Boom Is Still Real
The boom is still real: AI has gone mainstream.
The 2026 AI Index shows capital and adoption still at historic highs — but the striking part is that consumer value is growing for real.
FIG. 01 — The 2026 Picture
Source: Stanford AI Index 2026
Organizations Using AI
88%
of surveyed organizations use AI (2025)
Using GenAI
70%
use generative AI in at least one business function
US Private AI Investment
$285.9B
US 2025 private AI investment
US Consumer Value
$172B
annual US consumer value (early 2026, +54%)
53%
population adoption in three years — faster than PC or internet
Stanford AI Index 2026
1,953
AI companies newly funded in 2025
Stanford AI Index 2026
<10%
AI Agent deployment remains in the single digits across nearly every business function
Stanford AI Index 2026 · Economy
So What
Consumer value is real and growing rapidly. But large-scale agentification has not happened yet — adoption is real, transformation is still early.
02 — AI Capability Is Accelerating
Capability is accelerating: benchmarks approach saturation.
The 2026 models are no longer chatbots. GPT-5.6, Claude Sonnet 5, and Claude Opus 5 are built for complex work — and benchmarks nearly topped out within a year.
FIG. 02 — Capability Curve
Source: Stanford AI Index 2026 · Technical Performance
Notable Frontier Models (2025)
90%+
Over 90% of notable frontier models were produced by industry.
GPT-5.6
A frontier model for complex work: coding / research / science / cybersecurity / computer use / design. Launched 2026.07, Sol / Luna updated 2026.08.
OpenAI
Claude Sonnet 5
Positioned as far more agentic: planning / browser / terminal / tool use / autonomous execution.
Anthropic
Claude Opus 5
Near-Fable-5 intelligence at half the price ($5 / $25 per M token). Major gains on long-running agent work.
Anthropic · 2026.07.24
So What
“AI can’t really write complex code” no longer holds. Benchmark capability is approaching saturation — real-world reliability and economic integration are the harder problem now.
03 — From Assistant To Agent
AI is shifting from a "tool that answers questions" "to a "system that executes work.""
Anthropic’s Economic Index (2026.06) finds Claude usage shifting from conversational assistance toward long-running agentic tasks.
FIG. 03 — The Cadence Shift
Source: Anthropic Economic Index · June 2026
Past Conversational assistant
User
↓
Question
↓
Claude
↓
Answer
Now Agentic execution system
User
↓
Goal
↓
Claude Agent
↓
Tools · Code · Files · Iterations
↓
Result
2024
Answer questions
2025
Execute tasks
2026
Run agents
Next
Long-running autonomy
So What
The real question is no longer whether AI can act, but whether it can act reliably enough for economically important work.
04 — The Productivity Paradox
But productivity has become harder to measure.
AI productivity is not one number. It depends on model generation, task type, user experience, workflow, and measurement — and the evidence itself is becoming contradictory.
FIG. 04 — The Productivity Paradox
Sources: Science · METR · NBER
| Period | Evidence |
|---|---|
| 2023 | ChatGPT writing tasks: completion time -40%, quality +18%. The classic early-gen-AI result.Science · Noy & Zhang |
| 2025 | METR software engineering RCT: 16 experienced open-source developers, 246 tasks, using the AI coding tools of the time took 19% longer to complete tasks.METR · 2025 RCT |
| 2026 | METR update: developers increasingly refuse to work without AI, so experiments now suffer severe selection bias. Raw data hints at gains (-18% to -4%), but the authors stress this is only a lower bound on true gains; no reliable average uplift can be given.METR · Feb 2026 |
| 2026 | Anthropic survey: 86% of Claude users self-report working faster — but perception and objective measurement diverge; case evidence shows over-reliance can create new costs (e.g., Uber exhausting an AI budget in months).Anthropic Economic Index · 9,700 users |
| 2026 | 5,305 Claude Code transcripts: potential time savings of 1.5×–13×, but the authors explicitly warn this is exploratory, not a controlled productivity estimate.METR · transcript analysis |
Pattern worth noting (Stanford AI Index 2026): measured gains are clearest in structured, easily measurable tasks — and smallest where deeper reasoning is required.
So What
The evidence is contradictory: METR measured 19% longer completion times, while Management Science field experiments show ~26% more tasks completed, and 86% of users self-report gains. The honest conclusion: AI productivity is not one number — it depends on task, model, experience, and method. And don’t claim “Claude Code is 13× faster”; that is only the upper end of an exploratory analysis.
05 — AI Price Compression
The cost of intelligence keeps falling.
280× is the 2022–2024 story. Into 2026, frontier capability is still getting cheaper — this is genuine technological deflation.
FIG. 05 — AI Price Compression
Historical: Stanford AI Index 2025 · Current: OpenAI / Anthropic
Historical Deflation 2022 → 2024
280× cost decline
GPT-3.5-level inference: $20 → $0.07 per million tokens.
2026 Reality Frontier capability keeps getting cheaper
AI price deflation continues
GPT-5.6 Sol — GA 2026.07.09 ($5 / $30) → self price cut to $4 / $20 on 2026.08.22 (input -20% · output -33%), promo price guaranteed through 2026.11.21
↓
Claude Sonnet 5 — launched $2 / $10 (2026.06.30) → permanent pricing from 2026.08.10; planned September increase canceled
Industrial Revolution
Price of physical labor → commoditized
Internet
Price of information distribution → commoditized
AI
Marginal cost of cognitive production → falling, and still accelerating
So What
As frontier intelligence becomes increasingly affordable, competition moves from “who has intelligence” to who can turn intelligence into reliable business outcomes.
06 — The Bottleneck Has Moved
The bottleneck has moved: from generation to reliable execution.
Capability ≠ reliable economic autonomy. On the value chain, scarcity is moving upward.
Past
Intelligence
↓
Human Execution
↓
Value
Now
AI Intelligence
↓
Agentic Execution
↓
Human Verification
↓
Business Value
Next
AI Intelligence
↓
Reliable Autonomous Execution
↓
Business Value
Why we read it this way (Stanford AI Index 2026): measured gains are smallest precisely where tasks demand deeper reasoning — the point where reliable execution, not generation, is the binding constraint.
So What
Our reading: what used to be scarce was generation; what is increasingly scarce is verification and reliable execution. The real competition starts in the middle of the value chain.
07 — Building Is Cheap
Building is cheap now. Prototypes are no longer a moat.
One person + AI can ship a prototype fast. But a prototype is not a product, and a product is not a business.
Before AI
Idea
↓
Development team
↓
Capital
↓
Prototype
After AI
Idea
↓
AI Coding
↓
Prototype (faster)
08 — Distribution Is Still Scarce
Distribution is still scarce: the problem AI doesn't solve.
AI can write code, build demos, and run agents. But it does not automatically create users, trust, proprietary data, domain knowledge, or brand equity.
Users
User base
Trust
Trust
Data
Data
Domain
Industry knowledge
Brand
Brand
So What
A beautiful AI demo is not a business. Products with users, trust, and distribution are. Competition moves from “building it” to “being found, trusted, and delivered reliably.”
09 — Expertise Is Being Repriced
Expertise is being repriced.
AI won’t automatically make you an expert. It changes the leverage on expertise — amplifying experts while putting pressure on routine, low-context work.
Expert×AI×Automation
Expert × AI × Automation — not AI vs Human, but Expert × AI
Low value
"I can use Claude."
The tool is not a barrier. This sentence is approaching zero value.
High value
"I automated my CFD analysis pipeline with AI to improve engineering decisions."
Domain knowledge + workflow redesign + decision systems — assets that compound.
Early signal (Stanford AI Index 2026): the effect is already uneven on early-career software developers and hiring pipelines — routine, entry-level work is being repriced first.
So What
AI does not automatically create expertise. It changes the return on existing expertise. Your domain experience is the fulcrum AI leverages.
10 — The AI Bubble Question
Not "is AI a bubble," "but "are valuations running ahead of value.""
$285.9B in private investment, 1,953 newly funded companies, Google capex above $150B a year — enormous capital, but is economic value scaling at the same speed?
FIG. 06 — The Internet Precedent
NASDAQ Composite · 2000 → 2002
NASDAQ
-77%
5,048 (2000.03) → 1,139 (2002.10)
But the internet left behind
Search · E-commerce · Cloud · Digital Advertising
Bubbles kill companies without business models. They don’t kill technology.
The capital efficiency puzzle
US private AI investment ($285.9B) is about 23× China’s, yet the model-performance gap has narrowed to 2.7%. The return on enormous capital is unproven.
Stanford AI Index 2026
Analyst view · likely to wash out
Hype products · thin AI wrappers · AI apps with no users
Analyst view · likely to survive
Infrastructure · foundation models · applications solving real industry problems
So What
AI can be both a transformative technology and an overvalued market. The better question is not “Is AI a bubble?” but — are AI valuations pricing in future productivity before businesses can actually capture it?
11 — What Actually Compounds?
What, actually compounds?
Tools fade. Models upgrade. A few assets matter disproportionately.
Q1
Domain Knowledge
Understanding reality
Physics constraints, industry norms, how the real world actually works.
Q2
Judgment
Decision quality
Which result to trust? Which direction to invest in? Who is accountable?
Q3
Distribution
Reach
Users, trust, brand, network effects.
Q4
Ownership
Assets
Products, content, data, equity — compounding happens on what you own.
✕
Low value (getting cheaper)
- Simple writing
- Simple code
- Simple design
- Information tidying
✓
High value (getting more important)
- Judgment — decision-making
- Creation — from 0 to 1
- Industry experience — domain knowledge
- Ability to build systems — reliable execution
Domain Knowledge+AI Capability+Judgment+Ownership = Long-term Value
Domain Knowledge + AI Capability + Judgment + Ownership = Long-term Value
12 — Final Thesis
AI is not replacing value. It is repricing value.
AI is not replacing value. It is repricing value.
What Could Change My Mind
Clear macro evidence that AI adoption shows up in measured productivity within a year — rather than the messy, task-dependent results we see today.
Conclusion
The gold rush is ending.
The AI revolution is just beginning.
The gold rush was about access. The revolution is about leverage. They are not the same thing.
References
Evidence Sources
Core evidence draws on 2026 academic and industry data; community feedback is used only as qualitative observation, not statistical proof.
01
Stanford AI Index 2026 Adoption · investment · consumer value · SWE-bench — main report core
hai.stanford.edu/ai-index/2026
02
Stanford AI Index 2026 — Economy 88% adoption · 70% GenAI · $285.9B · $172B · US-China gap 2.7% · single-digit agent deployment core
hai.stanford.edu/ai-index/2026/economy
03
Stanford AI Index 2026 — Technical Performance SWE-bench 60% → ~100% · 90%+ frontier models from industry core
hai.stanford.edu/ai-index/2026/technical-performance
04
METR — Developer Productivity Experiment Update 2025 measured 19% longer completion times → 2026: selection effects prevent a reliable uplift core
metr.org · Feb 2026
05
METR — Claude Code Transcript Analysis 5,305 transcripts · potential 1.5×–13× savings · exploratory only core
metr.org · Feb 2026
06
Anthropic Economic Index — June 2026 Claude shifting from chat to long-running agentic sessions · 9,700-user survey (86% self-report faster) core
anthropic.com/economic-index
07
OpenAI — GPT-5.6 Sol GA 2026.07.09 ($5 / $30) · self price cut to $4 / $20 on 2026.08.22 (-20% / -33%) · promo guaranteed through 2026.11.21 core
openai.com/index/gpt-5-6
08
Anthropic — Claude Sonnet 5 Launched 2026.06.30 · $2 / $10 → permanent pricing from 2026.08.10, planned increase canceled core
anthropic.com/news
09
Anthropic — Claude Opus 5 Launched 2026.07.24 · near-Fable-5 intelligence at half price ($5 / $25) · long-running agent work core
anthropic.com/news
10
NBER — Generative AI at Work Historical productivity baseline · 5,179 customer support workers hist
nber.org/papers/w31161
11
Science — Experimental Evidence on Generative AI Writing tasks -40% time / +18% quality · Noy & Zhang hist
science.org/doi/10.1126/science.adh2586
12
Management Science — 2026 Three field experiments with software developers · ~26% more tasks completed in some conditions core
pubsonline.informs.org
13
Stanford AI Index 2025 Historical context: 280× cost decline (2022.11 → 2024.10) hist
hai.stanford.edu/ai-index/2025
14
METR — 2025 Developer RCT 16 experienced open-source developers · 246 tasks · 19% longer completion times hist
metr.org · Jul 2025
Open Questions
- Who captures the economic value as intelligence becomes abundant?
- How fast does AI adoption translate into measurable productivity?