Wednesday’s front page carried two stories that looked unrelated: a grocer cutting 3,000 corporate jobs without once blaming AI, and Google selling frontier intelligence at roughly a fifth of its rivals’ list price. They are the same story on two clocks — and the gap between those clocks is where the next few years of margins go.
🔍 THE BOTTOM LINE
The machine margin is changing address: the spread between what intelligence costs and what work sells for is moving from anyone who sells routine cognition at 2025 rates to whoever buys frontier capability at $2 per million tokens and resells it embedded in decisions, prices and store floors — as this morning’s Kroger, IMF, Argon and self-replication stories each showed one slice of.
The two schedules underneath Wednesday’s headlines
Take the four stories this site shipped this morning and line them up by speed. Kroger will cut 3,000 corporate roles over three years, framed by its New Zealand-born CEO as having “more layers, more handoffs and more overhead than we need” — gradual cuts, spread deliberately across 2027 to 2029. The IMF put numbers on the same wave: about 40 percent of jobs worldwide affected, roughly 60 percent in advanced economies, delivered as a “slow-moving tsunami” in which young workers stand in the surf zone. Gemini 4 Argon went public at $2/$10 per million tokens — introductory pricing The Verge calculates at roughly eighty percent below what OpenAI’s GPT-6 Astra and Anthropic’s Fable 5 charge. And Anthropic’s red team showed capable models nearly ready to copy themselves across $55 Android TV boxes, distributed inference at 0.39 tokens per second, no datacentre in sight.
Four stories, one schedule: the input cost of cognition is falling fast, while the organisations built on the old price rearrange slowly. Kroger’s three-year window is not caution — it is what the calendar of a 409,000-person workforce looks like when layered structure meets a capability curve. Georgieva’s “slow-moving” is doing the same work: the wave has already made landfall in the entry rungs; the visible restructuring arrives years later. The vanishing first rung this site traced on Monday is the leading edge of exactly that lag.
The pricing thread is the fastest clock of the four. Argon’s $2 is not an outlier — it is the visible mark of a price line that has been cut all year: DeepSeek’s discounting, Mistral’s intro rates, Cloudflare’s Clef-flash at $0.038 per million. What is new is that the price is being set by the company with the most vertically integrated stack — its own chips, its own datacentres, its own distribution in Search and Android — and by a player trying to buy back share after a year of trailing.
Who is acting like they can see it
Three players are positioned for the falling input price. Google is the obvious one: a price-led share grab the way it took cloud — buy the marginal developer at $2, double the price on 30 November once workflows have migrated, and keep successors (“Barium” and “Carbon”, reportedly already in internal testing per the pricing guide) behind it. The cache price at 95 percent off is the tell: Google is optimising for agentic workloads that re-read the same context, the usage pattern that grows fastest.
Kroger is the quieter exhibit. Whatever coverage credits the cuts to AI, the memo’s operative phrase is “look closely at where work is best fit to be done” — a sourcing decision, not a technology one — and the savings are directed at prices and store floors, not a lab. Foran is a margin-and-shelf-space operator using intelligence as a lever on cost, the way his trade always has. DeepSeek is making the same bet from the other side: a $400 million robot lab is a wager that when models are commoditised, the margin moves into the physical work the models direct — where open weights are the cheap input, not the product.
Anthropic sits paradoxically on both sides: it sells Fable 5.1 at $10/$50 while its own researchers spend the week showing that inference is about to become ambient on consumer junk. Read the self-replication paper as pricing research from the other end of the curve — the lab is defending the premium tier while openly documenting the world where model shards run on living-room hardware. The GitHub temperature check agrees that this direction is real: llama.cpp, the open project that puts inference on whatever hardware is nearby, sits at 131,073 stars and logged new commits within the last day — one of the most actively maintained codebases in open source, and it is not trying to sell anything. Wikipedia’s history of the project notes it began in March 2023 as one engineer’s Apple Silicon hobby; three and a half years later it is the substrate the entire edge-inference argument rests on.
Who is acting like the price is unchanged
Described by actions, not verdicts: the layered corporate structure Foran’s memo critiqued — every handoff is a place where purchased cognition now lands before a human touches it, and three-year restructuring windows price it as if the input cost were flat. Vendors and contractors with knowledge-work rates locked into multi-year renewals written in 2025 are in the same position, hedged by one honest caveat this site has had to make before: per-token price is not per-task price, and agent-driven workloads can burn more tokens even as tokens get cheaper — the software-postings rebound the Indeed data showed is the same two-sided coin, postings rising while the senior share of them grows. Budgets that assume $10/$50 lists hold forever are as wrong as budgets that assume the introductory $2 is permanent; the safe planning number is the direction of travel, which has pointed down through the whole of 2025 and 2026.
The wiki’s own history is the tension worth naming: OpenAI’s audited 2025 losses leaked earlier this year, the frontier labs are still spending at infrastructure-record rates, and the Microsoft-OpenAI relationship has already been rewritten once. List prices at $10/$50 do not cover what it actually costs to train and serve them; only the vertically integrated player can afford $2 as a loss-leader. That is the deep pattern — the price is being set by whoever’s balance sheet can survive it.
What it means for New Zealand
New Zealand buys frontier intelligence at US dollar list prices with a favourable-position tax on top: the exchange rate. Argon’s $2 introductory period that ends 30 November is the first genuinely affordable entry window for frontier-class models this year, and it is the moment for local agencies and developers to price build-versus-buy against the doubled rate they will actually pay from December. For workers, the Georgieva two-speed maps here precisely: MBIE’s modelling put 2 percent of NZ occupations at high automation risk but 36 percent at high augmentation potential, and the growth Georgieva cited — construction, data centres — is the physically present end, which is also where this month’s task-level research says the durable human role sits: licensed sign-off, physical action, trust. The local margin to capture is on that side, not in routine cognition resold at old rates. And the rung problem does not fix itself: the cheaper intelligence gets, the stronger the case for deliberately rebuilding the first rung — someone has to own the learning years.
FAQ
What happened to AI prices this week? Google moved Gemini 4 Argon, its frontier model, from a restricted cyber-defender cohort to general availability at $2/$10 per million tokens — roughly a fifth of what GPT-6 Astra and Claude Fable 5 charge per The Verge’s comparison — with prices doubling to $4/$20 from 1 December.
Did Kroger blame AI for its layoffs? No. CEO Greg Foran’s memo blamed “more layers, more handoffs and more overhead than we need,” while directing the company to adopt technology and look at “where work is best fit to be done.” AI is the enabler in the wording; the org chart is the culprit.
Why is the IMF calling AI a slow-moving tsunami? Kristalina Georgieva told the Bangkok annual meetings on 12 October that IMF research estimates about 40 percent of global jobs will be augmented, downgraded or eliminated — 60 percent in advanced economies — with the routine entry-level tasks young workers traditionally start on absorbing first.
Is cheaper AI good for New Zealand businesses? Yes for builders, complicated for everyone in knowledge services: API costs drop sharply (especially before 30 November), but the same drop compresses what routine digital work can be billed at. The advantage sits in physically present, regulated and trusted work that AI complements rather than replaces.
🔍 THE BOTTOM LINE
Whoever buys intelligence at $2 and sells it embedded in prices, products and decisions keeps the spread; whoever sells routine cognition at 2025 rates loses it slowly, over exactly the kind of three-year window Kroger just scheduled — which is why Georgieva called it a slow-moving tsunami rather than a crash. Watch 30 November: when the introductory price doubles, the floor Argon reset becomes the number every other lab has to live with.
📰 Sources
- AP News — Kroger to lay off 3,000 corporate employees over 3 years as it leans into AI (13 October 2026)
- Reuters, Fortune, CNBC, Bloomberg — IMF-World Bank Bangkok annual meetings coverage (12-13 October 2026)
- Google blog, The Verge, Reuters — Gemini 4 Argon general availability and pricing analysis (13-14 October 2026)
- Anthropic — AI models may be capable of self-replicating without access to a datacenter (13 October 2026); CNBC
- Wikipedia — llama.cpp entry and api.github.com repository metrics (retrieved 14 October 2026)
- Singularity.Kiwi — pricing guide data, updated 14 October 2026
