IssuesAllIssue 08 · 8 NovIssue 07 · 24 OctIssue 06 · 10 OctIssue 05 · 27 SepIssue 04 · 11 MayIssue 03 · 3 MayIssue 02 · 29 AprIssue 01 · 17 Apr
Rx AI Briefing

The Price of Thought

Issue 05 · Coverage 10–25 Sep 2026 · Issued 27 Sep 2026·14 min read

The cost of a given level of AI performance is falling about 47% a quarter — faster than computing, batteries, electricity or DNA sequencing ever did. Intelligence is becoming abundant. The constraint is moving away from producing an answer and towards absorbing what the answer changes.

Audio edition · The Price of Thought
0:009:47
This week · three headlines and one listen

Read separately they are a consumer app, an open-source pile-on and a pricing change. Read together they describe one shift: intelligence is becoming abundant, and the scarce thing is now the capacity to absorb it.

Recommended listen
Tobi Lütke on The Knowledge Project — 15 September · youtube.com/watch?v=G9P9D9hptq8 — the opportunity and the failure mode inside the same company. Shopify’s River, and Lütke’s “slop grenades”. See the closing section.
The Brief · intelligence is becoming abundant

“Since 2023, the cost of achieving the same level of AI performance has fallen by around 47% every quarter — roughly 13-fold each year.

  • Jan 2025 · o3 · past 25% on a hard mathematics benchmark · ~$0.55
  • ~18 months later · GPT-5.6 Luna · same threshold · ~$0.0015
  • Epoch puts that at a 377-fold reduction

Faster than computing, batteries, electricity or DNA sequencing.”

— Epoch AI, The plunging price of thought, September 2026. Published with the dataset and replication code.

The immediate conclusion is that intelligence is becoming cheaper. The more important conclusion is that it will therefore be used far more often — in far more places, by consumers, employees, software and machines.

Which moves the constraint. It shifts away from producing an answer, and towards absorbing the consequences: understanding what the answer changes, coordinating the response, and learning from what happens next.

The price of thought is collapsing. The value is moving to what a person or organisation can understand, absorb and do next.

Consumer AICommerceOperating Model
Story 01 · Muse

Customers gain a second brain

Meta shipped an agent that shops, books and fills in forms. The market did not reprice Meta. It repriced everyone whose advantage was the customer's effort.

Meta’s Muse is an early picture of what abundant intelligence looks like in everyday life: consumers delegate shopping, travel booking and form-filling. Within two weeks it recorded 2.8 million downloads and overtook ChatGPT as the top free app in the US and Canadian app stores.

The market reaction revealed the strategic issue more clearly than the product launch did. Meta and participating infrastructure companies rose. Shopify gained after enabling Muse checkout; PayPal, Stripe and Ticketmaster moved to connect. Amazon blocked it. And investors marked down companies whose customer access, comparison advantage or loyalty could be weakened by an agent acting on the consumer’s behalf.

Muse is not simply another digital assistant. It is the beginning of a second brain for the consumer — one that remembers preferences, searches continuously, compares without fatigue and increasingly acts.

That changes the balance of power. Customers who once tolerated complexity, inertia and imperfect information can now delegate the effort required to switch.

Figure 1 · rebuilt from the source data

The price of artificial thought has collapsed faster than that of any other transformative technology on record

How much cheaper a constant level of capability has become since each technology’s price decline began. Inflation-adjusted. Log scale — each gridline is a further order of magnitude. This is the curve underneath all three stories.

1×10×100×1k×1M×1B×1T×AI
2021–26
020406080
Years since the price decline began
  • Electricity1892–1973
  • Lithium batteries1991–2024
  • Compute1940–2001
  • DNA sequencing2001–2022
  • AI inference2021–2026
Chart rebuilt by Rx from the published dataset. Data: Epoch AI, “The plunging price of thought”, September 2026, used under CC BY. Underlying series: electricity, Hersh (1999) and Edison Electric Institute (1996); compute, Nordhaus (2001); lithium-ion cells, Our World in Data; DNA sequencing, National Human Genome Research Institute.
Figure 2 · same dataset

Compounded price decline per quarter, after inflation

Every technology on this list transformed an economy. None of them moved at this speed.

AI inference · 2023–26 · 12.7× a year
47.0%
DNA sequencing · 2001–22 · 1.84× a year
14.2%
Compute · 1940–2001 · 1.51× a year
9.9%
Lithium batteries · 1991–2024 · 1.16× a year
3.6%
Electricity · 1892–1973 · 1.05× a year
1.2%
AI inference prices are falling 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium-ion cells and 54× faster than electricity in the eighty years to 1973. Source as Figure 1.
The commercial test
Become easy for machines to buy from — without becoming invisible to customers.

Not another internal chatbot. A shared organisational mind.

If Muse is a second brain for an individual, the enterprise equivalent is not another internal chatbot. It is a trusted colleague through which employees and specialised agents can understand the company, collaborate across boundaries, move work forward and learn from outcomes.

Most enterprise assistants remain at the first stage. They retrieve a policy, summarise a meeting, draft an email. A real organisational agent would connect the customer described in CRM, the promise made in a contract, the invoices in finance, the service history in operations and the practical knowledge held by employees. It would understand the desired outcome, assemble the right context, prepare or take permitted actions, follow the work through, and fold the result back into organisational memory.

That is closer to a nervous system than a chatbot: it senses what is happening, connects relevant knowledge, coordinates a response, and learns from the consequence.

Enterprises are not there because the underlying technology is missing. They are not there because most companies have digitised activities without making the organisation legible to itself. Systems preserve transactions and documents, but often lose why a decision was made, which alternative was rejected, where an exception mattered and whether the eventual outcome was any good. The real operating model stays scattered across applications, teams, meetings and people’s memories.

A universal search box cannot fix that.

Enterprises need to build shared understanding journey by journey — starting where knowledge is fragmented, hand-offs are expensive and outcomes are observable. Complex complaints, sales-to-delivery, supplier onboarding and operational incidents are better starting points than indexing the entire company.

Observe
Watch the journey and make it legible before changing it.
Recommend
Suggest the next action, with the context that justifies it.
Prepare
Assemble the action, ready for a person to release.
Act
Take the action with approval, inside agreed boundaries.
Own
Handle the stable part of the work autonomously, with exceptions escalated.

Each completed journey should add reusable understanding about customers, products, policies, decisions and exceptions. Over time those connected journeys — not a single giant model — become the enterprise brain.

Board-level question
  • Consumers are gaining agents that understand their intentions. Do we have anything that understands how our business actually works — or only tools that answer questions about it?
AI ModelsOpen SourceArchitecture
Story 02 · Jev

Decisions become infrastructure

A model that returns a decision rather than prose went viral, and was reproduced in days. The interesting part is not the model. It is how fast the idea stopped being anyone's property.

Jev challenges the assumption that every AI task needs a large model composing a long answer. TypeSafe positions it as a System One model: it receives context and a structured question, then returns a choice, a score or a yes/no decision with a confidence value — rather than generating prose token by token.

This matters because a large share of the work inside software is not writing. It is deciding. Which request to route. Whether a case needs review. Which tool an agent should use. Whether an answer meets a threshold. What should happen next.

Compact decision models can make those choices faster and more cheaply, leaving larger models for ambiguity, synthesis and genuinely complex reasoning.

Jev launched on 15 September. Within days, open alternatives and replications appeared.

  • 15 Sep 2026Jev launches — closed weights, hosted, a typed decision instead of prose.
  • Within daysJulia 1 ships under Apache 2.0 — 144.3M parameters, runs on a CPU, and reported better results than the supplied Jev references on three of four pilot tasks.
  • The caveatIt fell significantly behind on the harder Banking77 test. The comparisons are early and narrow.
  • Late Sep 2026One ecosystem directory is already tracking 41 open models and codebases rebuilding the System One pattern.

The benchmark comparisons are early, narrow, and not the same thing as robust performance in enterprise use. But the direction is not in doubt: a useful architectural idea moved from proprietary novelty to a crowded open ecosystem in days, not years. Frontier labs are likely to absorb the pattern too — using smaller or specialised models to route work, check outputs and control more expensive ones.

Board-level question
  • If the model we are building on were free and universal tomorrow, what would still be ours — and is anyone accountable for growing it?
EconomicsEnterpriseMicrosoft
Story 03 · Microsoft

From seats to activity

Copilot was rebuilt around three surfaces. The commercial model underneath it is the more significant change: enterprise software moving from paying for access to paying for work performed.

Microsoft’s new Copilot brings Home, Code and Autopilot into one connected experience. Home combines conversation and Office work. Code lets employees create applications and workflows. Autopilot is a persistent agent that continues work in the cloud, monitors objectives and coordinates over time.

1

Home

Conversation and Office work in one place.

2

Code

Employees describe an application or workflow; Copilot builds it.

3

Autopilot

A persistent agent that keeps working, watches objectives and coordinates over time.

The commercial model is as significant as the product. Everyday Copilot use stays inside the user subscription; advanced capability draws on Copilot Credits. Microsoft describes this as a usage-based layer that complements subscriptions — cost depends on the complexity of the response, action or operation, and credits pool across the tenant. New Microsoft 365 Copilot Business purchases through cloud partners are due to have pay-as-you-go billing enabled by default from 2 November 2026.

This signals the evolution of enterprise software from paying primarily for access to paying for work performed. The licence does not disappear. Organisations may pay once to equip the employee, and again when the agent undertakes more substantial work.

Until now
You bought access, by headcount.
From here
You buy access and activity — which does not scale with headcount.

Large enterprises need to reprice around outcomes, not around seats:

  • Remove overlap. Retire AI tools and licences that provide the same capability twice.
  • Separate included from metered. Know precisely what the subscription covers and what attracts a usage charge.
  • Attribute variable spend to the business activity and the outcome that caused it.
  • Route by consequence. Cheap intelligence for routine work; expensive models only where extra capability changes the result.
  • Track the whole process cost — including human review, rework, delay and the downstream demand that gets created.
  • Renegotiate often. Capability, price and packaging are moving too fast for a static three-year assumption.

The governing measure is not cost per licence, per token or per interaction. It is cost per completed and useful outcome: a resolved case, a retained customer, a completed order, a faster decision, an avoided loss.

Microsoft is turning intelligence into a metered utility. Enterprises must learn to meter value just as carefully as vendors meter consumption.

Board-level question
  • We can see what our AI consumes. Can we see what it completed — and is anyone comparing the two?
The three, side by side

Different stories. The same shape.

Each of these developments has a first-order effect that sounds like a saving. Follow each of them out four steps and they arrive at the same place: a change in where advantage sits.

Figure 3 · the comparison

Three events, traced to four orders of consequence

Read down a column for the story. Read across a row for the argument.

Order of effectMusethe consumer agentJevthe decision modelCopilotthe meter
FirstSearch and purchase become easier for the consumer.Structured decisions become cheaper.A new variable cost appears.
SecondGreater price transparency, weaker habitual loyalty.Agent workflows become viable at far higher volumes.A licence estate built around headcount must govern machine activity.
ThirdAn external agent owns discovery, comparison and transaction — the provider becomes an interchangeable supplier.One general model gives way to an orchestrated portfolio: compact models for repetitive choices, large models for hard reasoning, deterministic code where no model is needed.Agent use can grow faster than the savings it produces.
FourthMarkets are redesigned around machine access: prices, availability, terms and service processes must be actionable by agents as well as people.Commoditisation. Model access alone cannot remain a durable enterprise advantage.Unit cost falls while total spend rises, because employees and agents consume far more.
The pattern is the point. In all three cases the first-order effect is that something becomes cheaper or easier, and the fourth-order effect is that a source of advantage moves somewhere else. Organisations that stop reading at the first order will buy the saving and miss the shift.
The human-speed bottleneck

The organisation looks more productive at the edge, and more congested at the centre.

These three stories point at the same tension. Consumers can delegate more work. Employees can produce more work. Agents can make more decisions. Vendors can charge for more machine activity. Yet most enterprises still absorb demand through human meetings, approvals, queues, hand-offs and manual exception handling.

AI added to the front of an unchanged process does not necessarily remove work. It creates more enquiries for service teams, more leads for sales, more analysis for managers, more changes for operations and more content for customers.

This is why employee productivity is an incomplete ambition. Helping every person produce 30% more can make the company worse if colleagues, controls and downstream systems cannot absorb the volume. The objective is not more individual output. It is more valuable work completed across the whole enterprise.

So the immediate leadership task is to redesign flow, not simply to deploy tools. Routine work should pass through with fewer hand-offs. People should concentrate on exceptions, relationships, competing interests and emerging problems. And capacity released at one stage should not be counted as a benefit if it merely creates cost at the next.

Understanding becomes the moat

Always one layer higher, one step ahead, and one more why.

It is tempting to conclude that judgment must always remain human. That is too static. Judgment itself can increasingly be delegated — once people understand a decision well enough to describe what good looks like, expose the trade-offs, supply feedback and recognise when conditions have changed.

The human role then moves up a layer: from producing an answer to evaluating it; from evaluating it to designing how evaluation happens; from designing the process to determining whether the organisation is solving the right problem. Once that higher layer stabilises, parts of it can be delegated too, and people move up again.

Figure 4 · the loop this publication is named after

The recursive cycle

Not a maturity model. A loop you re-enter at a higher level each time, which is why it does not terminate.

1
Understand the problem, its purpose, context and consequence.
2
Abstract the pattern, so it holds beyond a single case.
3
Delegate the stable part — to a person, a process or a machine.
4
Absorb the additional speed and volume that delegation creates.
5
Learn from the outcomes, and from the exceptions especially.
6
Move up to the next unresolved layer, and begin again.
and begin again, one layer higher
The scarce capability is not judgment on its own. It is the ability to build understanding continuously — to connect outcomes with causes, turn experience into reusable context, and ask one more “why” than the current system can answer.

Cheap intelligence will be available to every enterprise. Organisational understanding will not. Advantage will belong to companies that turn machine output into shared understanding, shared understanding into better orchestration, and better orchestration into the next generation of capability.

That is where the moat sits: with those who can curate and coordinate the next step, rather than merely react to what the technology produces. The distinctive asset is not a static knowledge base. It is a living feedback system that gets better at deciding what matters, which intelligence to apply, and what the organisation should learn next.

Recommended listen · Shopify's operating lesson

The recent Knowledge Project conversation with Shopify CEO Tobi Lütke is a useful companion to these events, because it shows both sides in one place.

Shopify’s internal agent, River, works in public Slack channels, can access code and data, run tests and open software changes. Lütke estimated that up to half of Shopify’s production pull requests now begin in a conversation with River. That is the enterprise colleague described in Story 01 — not a private answer box, but an agent embedded in shared work, where people and machines can see, challenge and learn from what happens.

Yet Lütke also describes “slop grenades”: AI-generated work passed to a colleague without the sender understanding, checking or improving it. The machine reduces the effort of production, and transfers the burden of interpretation and verification downstream.

When output is expensive, organisations worry about producing too little. When output becomes nearly free, the greater danger is producing more than the business can understand or absorb.

Shopify matters here not because it has resolved that tension, but because it is meeting it early. River shows the potential of shared organisational intelligence; slop grenades show what happens when production scales faster than understanding and accountability.

The next-generation enterprise will not win by keeping people in every decision, nor by removing them indiscriminately. It will win by moving people and machines together through successive layers of understanding — delegating what has become clear, learning from the consequence, and directing human attention towards what remains unresolved.

The price of thought is collapsing. The enduring premium is on understanding what deserves thought, absorbing what thought produces, and seeing the next move before it becomes obvious.

Issue compiled 27 September 2026, covering 10–25 September. Figures 1 and 2 were rebuilt by Rx from the dataset published with Epoch AI, “The plunging price of thought”, used under CC BY. Underlying series in Figure 1: electricity, Hersh (1999) and Edison Electric Institute (1996); compute, Nordhaus (2001); lithium-ion cells, Our World in Data; DNA sequencing, National Human Genome Research Institute. Other sources: Epoch AI on the o3 to GPT-5.6 Luna cost comparison; Reuters, “Meta’s Muse rekindles fears over winners and losers as personal AI agent emerges” (23 September 2026) and CNBC on the Amazon standoff; TypeSafe AI and the System One Models directory; Supersonic Labs and the Julia 1 model card; the awesome-jev ecosystem directory; Microsoft Partner Center September 2026 announcements, the Copilot Credits Guide (September 2026) and VentureBeat on the Copilot revamp; Tobi Lütke on The Knowledge Project with Shane Parrish, and reporting on “slop grenades” from Business Insider and Search Engine Journal.

The recursive edition

You have read the first degree.

The public issue traces three events out to their fourth-order effects. The recursive edition goes underneath them — at the assumptions the first read had to take on trust, what would have to be true for the argument to fail, and the question it did not reach.

One email to verify the address, then the issues. No sharing, and one click to stop.