← The stand
Cover of The Collection, Volume 1, Number 62: The Meter. Friday 30 October 2026, Melbourne. A watercolour of an old brass meter with a glass dial on a wooden desk, a long paper receipt curling off the desk edge, beside an open ledger whose pages are blank. Skip cover

Vol. 1  ·  No. 62  ·  Friday 30 October 2026  ·  Melbourne


The Collection

The Meter

Collected and edited by Newsletter World, for AK

Contents

Friday’s letters are all about the price of a machine answer, and they all reach the same awkward place. One annual report says the cost of a given score is falling about thirteenfold a year. A pricing letter says the seat licence is dying because the product got too efficient to bill by the head. A governance digest finds the money pooling at the sellers. A French lab puts its biggest model out with the rate card attached. And a newsroom spends a week with an AI coworker and decides it is not worth a review. The answer got cheaper. The meter kept running.

  1. iiiEditor’s LetterThe answer got cheaper. The meter kept running.03
  2. ivThe DiscountThirteen times cheaper a year, and the profit line sits still.04
  3. vThe SeatBilling by the head stops working when the product does the work.05
  4. viThe SellersPaying off for the people who sell it.06
  5. viiThe Rate CardOpen weights, with the price printed on the box.07
  6. viiiThe CoworkerHalf the price, and not worth a review.08
  7. ixStanding OrdersFour rules for this issue.09

03  ·  Editor’s Letter

The answer got cheaper. The meter kept running.

A meter does not care what you did with the electricity. It counts what went through the wire. If the price per unit falls and you leave every light on, the bill still arrives, and it may be bigger than last quarter’s.

That is where this week’s letters land on AI. The cost of reaching a fixed benchmark score is collapsing faster than electricity, computing or gene sequencing ever did. The two biggest labs’ combined revenue run rate has more than tripled, to about $105 billion. And the share of companies that can find any profit from AI in their books is about where it was twelve months ago.

None of that is a contradiction. An agent that does one job can make dozens of billed calls to do it. The labs are paid when the meter turns. Their customers are paid only once they change how the work is done, and that part has no discount. Read this issue as a bill, and look for the line where the savings were meant to show up.

The Collection · The Meter03

04  ·  The Discount

Thirteen times cheaper a year, and the profit line sits still

A reading of this year’s State of AI report, set against the survey numbers that ask what the buyers got for it.

Start with the discount, because it is real. Epoch AI’s figures, as the letter reports them, have the cost of reaching a fixed benchmark score falling about 47% a quarter since 2023. Compounded, that is roughly thirteen times cheaper every year. The fall is steepest just after a new top score, around 66% a quarter, and slows to about 32% a quarter two years on.

The worked example is GPQA Diamond, a graduate-level science test. Answering a question cost about 30 cents in early 2025. Some eighteen months later, GPT-5.6 Luna does it for $0.0004. That is about 725 times cheaper. On the letter’s comparison, AI’s costs have fallen 54 times faster than electricity’s did across the century to 1973, six times faster than computing’s, and about four times faster than DNA sequencing’s.

The labs get paid the moment their models are used. Their customers only see profit after they change how they work.

Then the bill. Nathan Benaich’s State of AI Report, the ninth, published on 8 October, has OpenAI and Anthropic together going from a combined revenue run rate of about $30 billion to about $105 billion by late summer 2026. On the buying side, McKinsey’s 2026 survey of 1,719 respondents found 80% saying AI made them more productive. Only 37% said it had added anything at all to earnings before interest and tax, about the same as in 2025. Around 6% count as high performers, with AI worth 5% or more of EBIT.

The letter’s explanation is the one on this paper’s cover. Cheaper requests do not mean smaller bills, because a single agent task can set off dozens of billed calls. Benaich calls the gap a human skill issue and imagines a Genius Bar for AI. It is fair to note that his own forecasts last year landed about 48% of the time, 31 of 59.

Two smaller numbers make the same point from the other end. One open model, GLM-5.1, went from 52.5% to 65.5% on a set of 100 SWE-bench Verified tasks with no change to the model, only to how it was set up. And inside Anthropic, the letter reports, Claude led 26% of research and development work in August, up from under 1% in February. The machine got cheaper. The arrangement around it is where the money is made or lost.

The Collection · The Meter04

05  ·  The Seat

Billing by the head stops working when the product does the work

The per-seat licence built the software industry. A pricing letter says the industry’s own products are now undoing it.

For twenty years the arithmetic was simple. More staff meant more seats, and more seats meant more revenue. The letter’s point is that AI turns that sum upside down. A vendor that charges per seat earns less precisely because its own product makes each customer need fewer people.

The numbers it gathers show the shift under way rather than finished. Kyle Poyar’s 2026 research, cited in the letter, puts hybrid pricing, a base fee plus usage, at 37% of companies, up from 25%. Investors asked to pick favour hybrid at 35% against 5% for pure seats. Yet 29% of companies above $150 million in annual recurring revenue still rely on seats, and 29% simply let the customer choose.

A vendor charging per seat earns less precisely because its own product becomes more efficient.

The buyers are not spending less, yet. Zylo’s 2026 figures have software spend per organisation up 8% to $55.7 million across about 305 apps. A Redpoint survey of 141 chief information officers, reported via SaaStr, found 45% paying for AI out of existing software budgets and 54% consolidating vendors. The money is moving between lines, not growing new ones.

The letter lays out the alternatives, charging by usage, by resolved ticket or by outcome, and gives the warning that matters most here. Know your cost per resolution, because inference is not a fixed cost. ZoomInfo’s chief executive, Henry Schuck, is quoted with the honest version: nobody knows where pricing ends up.

The Collection · The Meter05

06  ·  The Sellers

Paying off for the people who sell it

A weekly governance digest reads the earnings and the layoffs side by side.

The digest’s first line could have been this paper’s claim. Responsible AI, it says, is paying off for the sellers, not the buyers. Its evidence is the week’s results. Microsoft’s Azure grew 48%, and the company reported $18.6 billion in free cash flow, about $10 billion above expectations.

In the same week, the digest lists job cuts at Netflix, Meta, Salesforce and Groupon among others. It does not claim a single cause. It sets the two columns next to each other and lets the reader add them up.

Responsible AI is paying off for the sellers, not the buyers.

Two other items carry the meter metaphor into the physical world. CrowdStrike’s 2026 report on threats to the legal sector says attacks on law firms have doubled since 2023, with 455 legal organisations named on leak sites, up 39%.

And in Pennsylvania, Archbald Borough unanimously turned down the Wildcat Ridge data centre after nine hearings, about 26 hours of testimony and a 777-page application. The cheap answer needs a building, power and water somewhere. In Archbald, the borough said no.

The Collection · The Meter06

07  ·  The Rate Card

Open weights, with the price printed on the box

Mistral releases its largest model for anyone to download, and publishes what it will charge to run it for you.

Mistral Large 4 is out with open weights under a modified MIT licence. Anyone can take the model and run it on their own hardware, which puts the meter in their own building.

For those who would rather rent, the rate card is in the announcement. $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The model has three reasoning settings, low, medium and high, with high as the default. Mistral says it ranks first among open models on Cognition’s FrontierCode benchmark.

Cached input is billed at a tenth of fresh input. Whoever designs the prompts sets the bill.

The detail worth noticing is the gap between those numbers. Cached input costs a tenth of fresh input, and output costs three times input. High reasoning, the default, means more output. So the same model, on the same day, can cost very different amounts depending on how a team builds around it, which is the AI Corner’s point arriving as a price list.

One practical date for anyone already calling Mistral: the mistral-large-latest name switches to the new model on 12 November.

The Collection · The Meter07

08  ·  The Coworker

Half the price, and not worth a review

A newsroom that tests models for a living spends a week with a cheaper one built to use a computer.

GPT-6.1 Terra is priced at half of GPT-6.1 Sol and pitched at computer use, clicking and typing through software the way a person would. Every’s team found it quick at simple jobs. It could not set up an article in Every’s own content system.

The verdict from Dan Shipper, Every’s chief executive, was short. “We ran it but didn’t see enough of a difference” to give it the full Vibe Check treatment. Half the price was not enough to change the work.

When a model is good enough to force you to use a bad harness, it tells you something.

The team’s recommendation for editing went to a different model, Gemini 4 Argon, and Shipper’s line about it is the sharpest in the letter: when a model is good enough to force you to use a bad harness, it tells you something. The setup again. The same lesson the benchmark tinkerers learned with GLM-5.1, from the other side of the desk.

The week with the AI coworker itself ends on a warning any manager will recognise. It gets things wrong with total confidence. A cheaper hour of work is only cheaper if somebody does not have to spend another hour checking it.

The Collection · The Meter08

09  ·  Standing Orders

Four rules for this issue

  1. I

    Read the bill, not the price.

    A falling price per answer says nothing about what a job costs. Count the calls each task actually makes.

  2. II

    Know your cost per resolution.

    Inference is not a fixed cost. Price what you sell knowing what each finished job costs you to run.

  3. III

    Fix the setup before buying the model.

    The same model scored thirteen points higher with a better harness. Try that before the next upgrade.

  4. IV

    Find the line where the savings land.

    If nobody can point to it in the accounts, the discount went to the seller.